<\/h6>\r\nThis blog post is from Giovanni Donati<\/a>, Senior Consultant at MathWorks. <\/em>\r\n<\/h6>\r\n \r\n<\/h6>\r\nAt the 2025 edition of the SIA Simulation Num\u00e9rique<\/a> conference, taking place at the Renault Techo Center next to Paris, teams of engineering students from across France took on a cutting-edge challenge in simulation and digital modeling. Representing INSA Hauts-de-France, students Pierre Br\u00e9geon\u00a0<\/a>and Jo\u00e3o Marcos Souza Dias\u00a0<\/a>took on the challenge\u2014and did so in an unexpected way.\r\n<\/h6>\r\nAlthough both are mechanical engineering students\u00a0with no prior experience in artificial intelligence, they decided to approach the SIA Student Challenge through the lens of deep learning. With guidance from their academic mentors Fabien B\u00e9chet, Bertrand Lallemand, and Franck Massa, they dove head-first into a new discipline.\r\n<\/h6>\r\nTheir toolkit of choice? MATLAB<\/strong>\u00a0and Deep Learning Toolbox<\/strong>.\r\n<\/h6>\r\n \r\n<\/h6>\r\n
The Challenge: Faster Crash Simulations with Limited Data<\/strong><\/p>\r\nThe core task of the challenge was to estimate a Pareto front of mass vs. intrusion from a limited set of 60 high-fidelity crash simulations. Each simulation was defined by variations in six structural thickness parameters and outputting complex time-dependent displacement data across a thousand nodes. The dataset included the initial position of each node (X0, Y0, and Z0), the material thicknesses for each node (Ep1, Ep2, Ep3, Ep4, Ep5, and Ep6), as well as set of displacements for each node over 30 timesteps, representing the deformation of the structure over time.\r\n<\/h6>\r\n![\"\"](\"https:\/\/blogs.mathworks.com\/deep-learning\/files\/2025\/07\/Picture1.png\")
\r\n<\/h6>\r\nFigure 1:<\/strong> P-Beam structure for shock absorption<\/em>\r\n<\/h6>\r\nThe data used by the students for this challenge was made available by Renault, and it can be downloaded via the SIA website<\/a>\u00a0upon request. Renault also provided the students with tools for the verification of the model's final score on a new data set, unknown to the students. This ensured the fair verification of models in the competition.\r\n<\/h6>\r\nIn other words, the goal of the competition was to replicate the crash simulation behavior quickly and accurately, without the time or computational resources typically required for full finite element analysis<\/a>. Crash tests can be quite expensive, timewise and in loss of materials (crashed vehicles). Computationally efficient simulation allows achieving the best design for shock absorption structures faster and at a low cost.\r\n<\/h6>\r\nAs sponsors of the student challenge, MathWorks offered two tickets for the 24 Hours of Le Mans to the winning students!\r\n<\/h6>\r\n![\"\"](\"https:\/\/blogs.mathworks.com\/deep-learning\/files\/2025\/07\/JoaoAndPierre.jpg\")
\r\n<\/h6>\r\nJoao and Pierre wearing their MathWorks t-shirts at the 24 Hours of Le Mans!<\/em>\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Smart Strategies for a Complex Problem<\/strong><\/p>\r\nTo solve this problem, the INSA HDF team developed and compared two alternative approaches:\r\n\r\n \t- Full neural network approximation<\/li>\r\n \t
- Model reduction supported by meta-modeling<\/li>\r\n<\/ul>\r\n
![\"\"](\"https:\/\/blogs.mathworks.com\/deep-learning\/files\/2025\/07\/FirstApproach.png\")
\r\n<\/h6>\r\nFigure 2:<\/strong> First neural network architecture, used for the initial approach<\/em>\r\n<\/h6>\r\nThe first approach led the INSA team to a good accuracy score (85.2%) but slow simulation time (about 1 hour for training and prediction). Using this model as a baseline for comparison, they improved the model and ultimately obtained impressive results, leading them to win the challenge! Let's dig into the winning approach.\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Model Reduction Supported by Meta-Modeling<\/strong><\/p>\r\nTheir second and more advanced solution, integrated a multiple-stage model reduction:\r\n<\/h6>\r\n\r\n \t- They applied Dynamic Model Decomposition (DMD) on each simulation, keeping only 15 key modes.<\/li>\r\n \t
- They used Singular Value Decomposition (SVD) to create a global reduced basis (300 modes).\r\n
<\/h6>\r\n![\"\"](\"https:\/\/blogs.mathworks.com\/deep-learning\/files\/2025\/07\/SVDComponents.png\")
\r\n<\/h6>\r\nFigure 3:<\/strong> Mode energies of SVD components<\/em><\/li>\r\n \t- They approximated the reduced coefficients by using a shallow neural network.
![\"\"](\"https:\/\/blogs.mathworks.com\/deep-learning\/files\/2025\/07\/WinningApproach.png\")
\r\n<\/h6>\r\nFigure 4:<\/strong> Neural network architecture of the winning solution<\/em><\/li>\r\n \t- They reconstructed the physical displacements.<\/li>\r\n<\/ol>\r\nThe full workflow is represented using the following principle scheme:\r\n
<\/h6>\r\n![\"\"](\"https:\/\/blogs.mathworks.com\/deep-learning\/files\/2025\/07\/SchemePrinciple.png\")
\r\n<\/h6>\r\nFigure 5:<\/strong> Principle scheme of the training and prediction process<\/em>\r\n<\/h6>\r\nBy adding extra data preprocessing steps, their new strategy achieved a slightly better accuracy and an impressive simulation speed-up compared to their first approach. They achieved:\r\n<\/h6>\r\n\r\n \t- Accuracy score: 86.2%<\/strong><\/li>\r\n \t
- Model Training and Prediction time: 150 seconds<\/strong><\/li>\r\n<\/ul>\r\nThe speed-up is 168x<\/strong>\u00a0compared to the first full machine learning trial!\r\n
<\/h6>\r\n![\"\"](\"https:\/\/blogs.mathworks.com\/deep-learning\/files\/2025\/07\/ScoreVisualization.png\")
\r\n<\/h6>\r\nFigure 6:<\/strong> Score and Pareto Front visualization provided by Renault for evaluating the quality of the models<\/em>\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Why MATLAB?<\/strong><\/p>\r\nFor students new to AI and data science, MATLAB provided a smooth on-ramp. Its intuitive syntax, built-in visualization, and integrated deep learning modeling\u00a0allowed the team to:\r\n<\/h6>\r\n\r\n \t- Experiment quickly with model architectures.<\/li>\r\n \t
- Easily visualize and analyze results.<\/li>\r\n \t
- Focus on their core engineering goals without losing time with figuring out the right tools.<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Looking Ahead<\/strong><\/p>\r\nThe team's work demonstrated the effectiveness of hybrid AI \/ physics-based approaches<\/a> in accelerating engineering simulations. Their strategy is not only technically elegant, but also directly applicable to real-word industrial settings where simulation time is a bottleneck.\r\n<\/h6>\r\nFuture improvements could include:\r\n<\/h6>\r\n\r\n \t- Improving the existing neural network<\/li>\r\n \t
- Exploring other decomposition techniques<\/li>\r\n \t
- Exploring other neural network architectures<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Final Thoughts<\/strong><\/p>\r\nThe INSA HDF team's contribution to the SIA Student Challenge highlights how emerging engineers can blend machine learning and numerical methods to push the boundaries of simulation science.\r\n<\/h6>\r\nCongratulations to Pierre, Joao, and their mentors for this outstanding achievement!\r\n<\/h6>\r\nWant to learn more about model reduction and AI engineering? Check out resources on reduced-order modeling and dive into deep learning with MATLAB.\r\n<\/h6>\r\nResources:<\/strong>\r\n<\/h6>\r\n\r\n \t- Finite Element Analysis with MATLAB<\/a><\/li>\r\n \t
- Physics-Informed Neural Networks<\/a><\/li>\r\n \t
- Reduced Order Modeling with MATLAB<\/a><\/li>\r\n<\/ul>\r\n
<\/h6>","protected":false},"excerpt":{"rendered":"<\/div>\r\nHighlights from the SIA Student Challenge - INSA Hauts-de-France Team\r\n\r\n\r\nThis blog post is from Giovanni Donati, Senior Consultant at MathWorks. \r\n\r\n \r\n\r\nAt the 2025 edition of the SIA... read more >><\/a><\/p>","protected":false},"author":194,"featured_media":18071,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[9],"tags":[],"_links":{"self":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056"}],"collection":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/users\/194"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/comments?post=18056"}],"version-history":[{"count":46,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions"}],"predecessor-version":[{"id":18236,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions\/18236"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media\/18071"}],"wp:attachment":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media?parent=18056"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/categories?post=18056"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/tags?post=18056"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
<\/h6>\r\nAt the 2025 edition of the SIA Simulation Num\u00e9rique<\/a> conference, taking place at the Renault Techo Center next to Paris, teams of engineering students from across France took on a cutting-edge challenge in simulation and digital modeling. Representing INSA Hauts-de-France, students Pierre Br\u00e9geon\u00a0<\/a>and Jo\u00e3o Marcos Souza Dias\u00a0<\/a>took on the challenge\u2014and did so in an unexpected way.\r\n<\/h6>\r\nAlthough both are mechanical engineering students\u00a0with no prior experience in artificial intelligence, they decided to approach the SIA Student Challenge through the lens of deep learning. With guidance from their academic mentors Fabien B\u00e9chet, Bertrand Lallemand, and Franck Massa, they dove head-first into a new discipline.\r\n<\/h6>\r\nTheir toolkit of choice? MATLAB<\/strong>\u00a0and Deep Learning Toolbox<\/strong>.\r\n<\/h6>\r\n \r\n<\/h6>\r\n
The Challenge: Faster Crash Simulations with Limited Data<\/strong><\/p>\r\nThe core task of the challenge was to estimate a Pareto front of mass vs. intrusion from a limited set of 60 high-fidelity crash simulations. Each simulation was defined by variations in six structural thickness parameters and outputting complex time-dependent displacement data across a thousand nodes. The dataset included the initial position of each node (X0, Y0, and Z0), the material thicknesses for each node (Ep1, Ep2, Ep3, Ep4, Ep5, and Ep6), as well as set of displacements for each node over 30 timesteps, representing the deformation of the structure over time.\r\n<\/h6>\r\n\r\n<\/h6>\r\nFigure 1:<\/strong> P-Beam structure for shock absorption<\/em>\r\n<\/h6>\r\nThe data used by the students for this challenge was made available by Renault, and it can be downloaded via the SIA website<\/a>\u00a0upon request. Renault also provided the students with tools for the verification of the model's final score on a new data set, unknown to the students. This ensured the fair verification of models in the competition.\r\n<\/h6>\r\nIn other words, the goal of the competition was to replicate the crash simulation behavior quickly and accurately, without the time or computational resources typically required for full finite element analysis<\/a>. Crash tests can be quite expensive, timewise and in loss of materials (crashed vehicles). Computationally efficient simulation allows achieving the best design for shock absorption structures faster and at a low cost.\r\n<\/h6>\r\nAs sponsors of the student challenge, MathWorks offered two tickets for the 24 Hours of Le Mans to the winning students!\r\n<\/h6>\r\n\r\n<\/h6>\r\nJoao and Pierre wearing their MathWorks t-shirts at the 24 Hours of Le Mans!<\/em>\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Smart Strategies for a Complex Problem<\/strong><\/p>\r\nTo solve this problem, the INSA HDF team developed and compared two alternative approaches:\r\n\r\n \t- Full neural network approximation<\/li>\r\n \t
- Model reduction supported by meta-modeling<\/li>\r\n<\/ul>\r\n\r\n
<\/h6>\r\nFigure 2:<\/strong> First neural network architecture, used for the initial approach<\/em>\r\n<\/h6>\r\nThe first approach led the INSA team to a good accuracy score (85.2%) but slow simulation time (about 1 hour for training and prediction). Using this model as a baseline for comparison, they improved the model and ultimately obtained impressive results, leading them to win the challenge! Let's dig into the winning approach.\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Model Reduction Supported by Meta-Modeling<\/strong><\/p>\r\nTheir second and more advanced solution, integrated a multiple-stage model reduction:\r\n<\/h6>\r\n\r\n \t- They applied Dynamic Model Decomposition (DMD) on each simulation, keeping only 15 key modes.<\/li>\r\n \t
- They used Singular Value Decomposition (SVD) to create a global reduced basis (300 modes).\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 3:<\/strong> Mode energies of SVD components<\/em><\/li>\r\n \t- They approximated the reduced coefficients by using a shallow neural network.\r\n
<\/h6>\r\nFigure 4:<\/strong> Neural network architecture of the winning solution<\/em><\/li>\r\n \t- They reconstructed the physical displacements.<\/li>\r\n<\/ol>\r\nThe full workflow is represented using the following principle scheme:\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 5:<\/strong> Principle scheme of the training and prediction process<\/em>\r\n<\/h6>\r\nBy adding extra data preprocessing steps, their new strategy achieved a slightly better accuracy and an impressive simulation speed-up compared to their first approach. They achieved:\r\n<\/h6>\r\n\r\n \t- Accuracy score: 86.2%<\/strong><\/li>\r\n \t
- Model Training and Prediction time: 150 seconds<\/strong><\/li>\r\n<\/ul>\r\nThe speed-up is 168x<\/strong>\u00a0compared to the first full machine learning trial!\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 6:<\/strong> Score and Pareto Front visualization provided by Renault for evaluating the quality of the models<\/em>\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Why MATLAB?<\/strong><\/p>\r\nFor students new to AI and data science, MATLAB provided a smooth on-ramp. Its intuitive syntax, built-in visualization, and integrated deep learning modeling\u00a0allowed the team to:\r\n<\/h6>\r\n\r\n \t- Experiment quickly with model architectures.<\/li>\r\n \t
- Easily visualize and analyze results.<\/li>\r\n \t
- Focus on their core engineering goals without losing time with figuring out the right tools.<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Looking Ahead<\/strong><\/p>\r\nThe team's work demonstrated the effectiveness of hybrid AI \/ physics-based approaches<\/a> in accelerating engineering simulations. Their strategy is not only technically elegant, but also directly applicable to real-word industrial settings where simulation time is a bottleneck.\r\n<\/h6>\r\nFuture improvements could include:\r\n<\/h6>\r\n\r\n \t- Improving the existing neural network<\/li>\r\n \t
- Exploring other decomposition techniques<\/li>\r\n \t
- Exploring other neural network architectures<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Final Thoughts<\/strong><\/p>\r\nThe INSA HDF team's contribution to the SIA Student Challenge highlights how emerging engineers can blend machine learning and numerical methods to push the boundaries of simulation science.\r\n<\/h6>\r\nCongratulations to Pierre, Joao, and their mentors for this outstanding achievement!\r\n<\/h6>\r\nWant to learn more about model reduction and AI engineering? Check out resources on reduced-order modeling and dive into deep learning with MATLAB.\r\n<\/h6>\r\nResources:<\/strong>\r\n<\/h6>\r\n\r\n \t- Finite Element Analysis with MATLAB<\/a><\/li>\r\n \t
- Physics-Informed Neural Networks<\/a><\/li>\r\n \t
- Reduced Order Modeling with MATLAB<\/a><\/li>\r\n<\/ul>\r\n
<\/h6>","protected":false},"excerpt":{"rendered":"<\/div>\r\nHighlights from the SIA Student Challenge - INSA Hauts-de-France Team\r\n\r\n\r\nThis blog post is from Giovanni Donati, Senior Consultant at MathWorks. \r\n\r\n \r\n\r\nAt the 2025 edition of the SIA... read more >><\/a><\/p>","protected":false},"author":194,"featured_media":18071,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[9],"tags":[],"_links":{"self":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056"}],"collection":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/users\/194"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/comments?post=18056"}],"version-history":[{"count":46,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions"}],"predecessor-version":[{"id":18236,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions\/18236"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media\/18071"}],"wp:attachment":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media?parent=18056"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/categories?post=18056"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/tags?post=18056"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
<\/h6>\r\nTheir toolkit of choice? MATLAB<\/strong>\u00a0and Deep Learning Toolbox<\/strong>.\r\n<\/h6>\r\n \r\n<\/h6>\r\n
The Challenge: Faster Crash Simulations with Limited Data<\/strong><\/p>\r\nThe core task of the challenge was to estimate a Pareto front of mass vs. intrusion from a limited set of 60 high-fidelity crash simulations. Each simulation was defined by variations in six structural thickness parameters and outputting complex time-dependent displacement data across a thousand nodes. The dataset included the initial position of each node (X0, Y0, and Z0), the material thicknesses for each node (Ep1, Ep2, Ep3, Ep4, Ep5, and Ep6), as well as set of displacements for each node over 30 timesteps, representing the deformation of the structure over time.\r\n<\/h6>\r\n\r\n<\/h6>\r\nFigure 1:<\/strong> P-Beam structure for shock absorption<\/em>\r\n<\/h6>\r\nThe data used by the students for this challenge was made available by Renault, and it can be downloaded via the SIA website<\/a>\u00a0upon request. Renault also provided the students with tools for the verification of the model's final score on a new data set, unknown to the students. This ensured the fair verification of models in the competition.\r\n<\/h6>\r\nIn other words, the goal of the competition was to replicate the crash simulation behavior quickly and accurately, without the time or computational resources typically required for full finite element analysis<\/a>. Crash tests can be quite expensive, timewise and in loss of materials (crashed vehicles). Computationally efficient simulation allows achieving the best design for shock absorption structures faster and at a low cost.\r\n<\/h6>\r\nAs sponsors of the student challenge, MathWorks offered two tickets for the 24 Hours of Le Mans to the winning students!\r\n<\/h6>\r\n\r\n<\/h6>\r\nJoao and Pierre wearing their MathWorks t-shirts at the 24 Hours of Le Mans!<\/em>\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Smart Strategies for a Complex Problem<\/strong><\/p>\r\nTo solve this problem, the INSA HDF team developed and compared two alternative approaches:\r\n\r\n \t- Full neural network approximation<\/li>\r\n \t
- Model reduction supported by meta-modeling<\/li>\r\n<\/ul>\r\n\r\n
<\/h6>\r\nFigure 2:<\/strong> First neural network architecture, used for the initial approach<\/em>\r\n<\/h6>\r\nThe first approach led the INSA team to a good accuracy score (85.2%) but slow simulation time (about 1 hour for training and prediction). Using this model as a baseline for comparison, they improved the model and ultimately obtained impressive results, leading them to win the challenge! Let's dig into the winning approach.\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Model Reduction Supported by Meta-Modeling<\/strong><\/p>\r\nTheir second and more advanced solution, integrated a multiple-stage model reduction:\r\n<\/h6>\r\n\r\n \t- They applied Dynamic Model Decomposition (DMD) on each simulation, keeping only 15 key modes.<\/li>\r\n \t
- They used Singular Value Decomposition (SVD) to create a global reduced basis (300 modes).\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 3:<\/strong> Mode energies of SVD components<\/em><\/li>\r\n \t- They approximated the reduced coefficients by using a shallow neural network.\r\n
<\/h6>\r\nFigure 4:<\/strong> Neural network architecture of the winning solution<\/em><\/li>\r\n \t- They reconstructed the physical displacements.<\/li>\r\n<\/ol>\r\nThe full workflow is represented using the following principle scheme:\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 5:<\/strong> Principle scheme of the training and prediction process<\/em>\r\n<\/h6>\r\nBy adding extra data preprocessing steps, their new strategy achieved a slightly better accuracy and an impressive simulation speed-up compared to their first approach. They achieved:\r\n<\/h6>\r\n\r\n \t- Accuracy score: 86.2%<\/strong><\/li>\r\n \t
- Model Training and Prediction time: 150 seconds<\/strong><\/li>\r\n<\/ul>\r\nThe speed-up is 168x<\/strong>\u00a0compared to the first full machine learning trial!\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 6:<\/strong> Score and Pareto Front visualization provided by Renault for evaluating the quality of the models<\/em>\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Why MATLAB?<\/strong><\/p>\r\nFor students new to AI and data science, MATLAB provided a smooth on-ramp. Its intuitive syntax, built-in visualization, and integrated deep learning modeling\u00a0allowed the team to:\r\n<\/h6>\r\n\r\n \t- Experiment quickly with model architectures.<\/li>\r\n \t
- Easily visualize and analyze results.<\/li>\r\n \t
- Focus on their core engineering goals without losing time with figuring out the right tools.<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Looking Ahead<\/strong><\/p>\r\nThe team's work demonstrated the effectiveness of hybrid AI \/ physics-based approaches<\/a> in accelerating engineering simulations. Their strategy is not only technically elegant, but also directly applicable to real-word industrial settings where simulation time is a bottleneck.\r\n<\/h6>\r\nFuture improvements could include:\r\n<\/h6>\r\n\r\n \t- Improving the existing neural network<\/li>\r\n \t
- Exploring other decomposition techniques<\/li>\r\n \t
- Exploring other neural network architectures<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Final Thoughts<\/strong><\/p>\r\nThe INSA HDF team's contribution to the SIA Student Challenge highlights how emerging engineers can blend machine learning and numerical methods to push the boundaries of simulation science.\r\n<\/h6>\r\nCongratulations to Pierre, Joao, and their mentors for this outstanding achievement!\r\n<\/h6>\r\nWant to learn more about model reduction and AI engineering? Check out resources on reduced-order modeling and dive into deep learning with MATLAB.\r\n<\/h6>\r\nResources:<\/strong>\r\n<\/h6>\r\n\r\n \t- Finite Element Analysis with MATLAB<\/a><\/li>\r\n \t
- Physics-Informed Neural Networks<\/a><\/li>\r\n \t
- Reduced Order Modeling with MATLAB<\/a><\/li>\r\n<\/ul>\r\n
<\/h6>","protected":false},"excerpt":{"rendered":"<\/div>\r\nHighlights from the SIA Student Challenge - INSA Hauts-de-France Team\r\n\r\n\r\nThis blog post is from Giovanni Donati, Senior Consultant at MathWorks. \r\n\r\n \r\n\r\nAt the 2025 edition of the SIA... read more >><\/a><\/p>","protected":false},"author":194,"featured_media":18071,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[9],"tags":[],"_links":{"self":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056"}],"collection":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/users\/194"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/comments?post=18056"}],"version-history":[{"count":46,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions"}],"predecessor-version":[{"id":18236,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions\/18236"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media\/18071"}],"wp:attachment":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media?parent=18056"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/categories?post=18056"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/tags?post=18056"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
<\/h6>\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 1:<\/strong> P-Beam structure for shock absorption<\/em>\r\n<\/h6>\r\nThe data used by the students for this challenge was made available by Renault, and it can be downloaded via the SIA website<\/a>\u00a0upon request. Renault also provided the students with tools for the verification of the model's final score on a new data set, unknown to the students. This ensured the fair verification of models in the competition.\r\n<\/h6>\r\nIn other words, the goal of the competition was to replicate the crash simulation behavior quickly and accurately, without the time or computational resources typically required for full finite element analysis<\/a>. Crash tests can be quite expensive, timewise and in loss of materials (crashed vehicles). Computationally efficient simulation allows achieving the best design for shock absorption structures faster and at a low cost.\r\n<\/h6>\r\nAs sponsors of the student challenge, MathWorks offered two tickets for the 24 Hours of Le Mans to the winning students!\r\n<\/h6>\r\n\r\n<\/h6>\r\nJoao and Pierre wearing their MathWorks t-shirts at the 24 Hours of Le Mans!<\/em>\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Smart Strategies for a Complex Problem<\/strong><\/p>\r\nTo solve this problem, the INSA HDF team developed and compared two alternative approaches:\r\n\r\n \t- Full neural network approximation<\/li>\r\n \t
- Model reduction supported by meta-modeling<\/li>\r\n<\/ul>\r\n\r\n
<\/h6>\r\nFigure 2:<\/strong> First neural network architecture, used for the initial approach<\/em>\r\n<\/h6>\r\nThe first approach led the INSA team to a good accuracy score (85.2%) but slow simulation time (about 1 hour for training and prediction). Using this model as a baseline for comparison, they improved the model and ultimately obtained impressive results, leading them to win the challenge! Let's dig into the winning approach.\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Model Reduction Supported by Meta-Modeling<\/strong><\/p>\r\nTheir second and more advanced solution, integrated a multiple-stage model reduction:\r\n<\/h6>\r\n\r\n \t- They applied Dynamic Model Decomposition (DMD) on each simulation, keeping only 15 key modes.<\/li>\r\n \t
- They used Singular Value Decomposition (SVD) to create a global reduced basis (300 modes).\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 3:<\/strong> Mode energies of SVD components<\/em><\/li>\r\n \t- They approximated the reduced coefficients by using a shallow neural network.\r\n
<\/h6>\r\nFigure 4:<\/strong> Neural network architecture of the winning solution<\/em><\/li>\r\n \t- They reconstructed the physical displacements.<\/li>\r\n<\/ol>\r\nThe full workflow is represented using the following principle scheme:\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 5:<\/strong> Principle scheme of the training and prediction process<\/em>\r\n<\/h6>\r\nBy adding extra data preprocessing steps, their new strategy achieved a slightly better accuracy and an impressive simulation speed-up compared to their first approach. They achieved:\r\n<\/h6>\r\n\r\n \t- Accuracy score: 86.2%<\/strong><\/li>\r\n \t
- Model Training and Prediction time: 150 seconds<\/strong><\/li>\r\n<\/ul>\r\nThe speed-up is 168x<\/strong>\u00a0compared to the first full machine learning trial!\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 6:<\/strong> Score and Pareto Front visualization provided by Renault for evaluating the quality of the models<\/em>\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Why MATLAB?<\/strong><\/p>\r\nFor students new to AI and data science, MATLAB provided a smooth on-ramp. Its intuitive syntax, built-in visualization, and integrated deep learning modeling\u00a0allowed the team to:\r\n<\/h6>\r\n\r\n \t- Experiment quickly with model architectures.<\/li>\r\n \t
- Easily visualize and analyze results.<\/li>\r\n \t
- Focus on their core engineering goals without losing time with figuring out the right tools.<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Looking Ahead<\/strong><\/p>\r\nThe team's work demonstrated the effectiveness of hybrid AI \/ physics-based approaches<\/a> in accelerating engineering simulations. Their strategy is not only technically elegant, but also directly applicable to real-word industrial settings where simulation time is a bottleneck.\r\n<\/h6>\r\nFuture improvements could include:\r\n<\/h6>\r\n\r\n \t- Improving the existing neural network<\/li>\r\n \t
- Exploring other decomposition techniques<\/li>\r\n \t
- Exploring other neural network architectures<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Final Thoughts<\/strong><\/p>\r\nThe INSA HDF team's contribution to the SIA Student Challenge highlights how emerging engineers can blend machine learning and numerical methods to push the boundaries of simulation science.\r\n<\/h6>\r\nCongratulations to Pierre, Joao, and their mentors for this outstanding achievement!\r\n<\/h6>\r\nWant to learn more about model reduction and AI engineering? Check out resources on reduced-order modeling and dive into deep learning with MATLAB.\r\n<\/h6>\r\nResources:<\/strong>\r\n<\/h6>\r\n\r\n \t- Finite Element Analysis with MATLAB<\/a><\/li>\r\n \t
- Physics-Informed Neural Networks<\/a><\/li>\r\n \t
- Reduced Order Modeling with MATLAB<\/a><\/li>\r\n<\/ul>\r\n
<\/h6>","protected":false},"excerpt":{"rendered":"<\/div>\r\nHighlights from the SIA Student Challenge - INSA Hauts-de-France Team\r\n\r\n\r\nThis blog post is from Giovanni Donati, Senior Consultant at MathWorks. \r\n\r\n \r\n\r\nAt the 2025 edition of the SIA... read more >><\/a><\/p>","protected":false},"author":194,"featured_media":18071,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[9],"tags":[],"_links":{"self":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056"}],"collection":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/users\/194"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/comments?post=18056"}],"version-history":[{"count":46,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions"}],"predecessor-version":[{"id":18236,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions\/18236"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media\/18071"}],"wp:attachment":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media?parent=18056"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/categories?post=18056"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/tags?post=18056"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
<\/h6>\r\nThe data used by the students for this challenge was made available by Renault, and it can be downloaded via the SIA website<\/a>\u00a0upon request. Renault also provided the students with tools for the verification of the model's final score on a new data set, unknown to the students. This ensured the fair verification of models in the competition.\r\n<\/h6>\r\nIn other words, the goal of the competition was to replicate the crash simulation behavior quickly and accurately, without the time or computational resources typically required for full finite element analysis<\/a>. Crash tests can be quite expensive, timewise and in loss of materials (crashed vehicles). Computationally efficient simulation allows achieving the best design for shock absorption structures faster and at a low cost.\r\n<\/h6>\r\nAs sponsors of the student challenge, MathWorks offered two tickets for the 24 Hours of Le Mans to the winning students!\r\n<\/h6>\r\n\r\n<\/h6>\r\nJoao and Pierre wearing their MathWorks t-shirts at the 24 Hours of Le Mans!<\/em>\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Smart Strategies for a Complex Problem<\/strong><\/p>\r\nTo solve this problem, the INSA HDF team developed and compared two alternative approaches:\r\n\r\n \t- Full neural network approximation<\/li>\r\n \t
- Model reduction supported by meta-modeling<\/li>\r\n<\/ul>\r\n\r\n
<\/h6>\r\nFigure 2:<\/strong> First neural network architecture, used for the initial approach<\/em>\r\n<\/h6>\r\nThe first approach led the INSA team to a good accuracy score (85.2%) but slow simulation time (about 1 hour for training and prediction). Using this model as a baseline for comparison, they improved the model and ultimately obtained impressive results, leading them to win the challenge! Let's dig into the winning approach.\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Model Reduction Supported by Meta-Modeling<\/strong><\/p>\r\nTheir second and more advanced solution, integrated a multiple-stage model reduction:\r\n<\/h6>\r\n\r\n \t- They applied Dynamic Model Decomposition (DMD) on each simulation, keeping only 15 key modes.<\/li>\r\n \t
- They used Singular Value Decomposition (SVD) to create a global reduced basis (300 modes).\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 3:<\/strong> Mode energies of SVD components<\/em><\/li>\r\n \t- They approximated the reduced coefficients by using a shallow neural network.\r\n
<\/h6>\r\nFigure 4:<\/strong> Neural network architecture of the winning solution<\/em><\/li>\r\n \t- They reconstructed the physical displacements.<\/li>\r\n<\/ol>\r\nThe full workflow is represented using the following principle scheme:\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 5:<\/strong> Principle scheme of the training and prediction process<\/em>\r\n<\/h6>\r\nBy adding extra data preprocessing steps, their new strategy achieved a slightly better accuracy and an impressive simulation speed-up compared to their first approach. They achieved:\r\n<\/h6>\r\n\r\n \t- Accuracy score: 86.2%<\/strong><\/li>\r\n \t
- Model Training and Prediction time: 150 seconds<\/strong><\/li>\r\n<\/ul>\r\nThe speed-up is 168x<\/strong>\u00a0compared to the first full machine learning trial!\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 6:<\/strong> Score and Pareto Front visualization provided by Renault for evaluating the quality of the models<\/em>\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Why MATLAB?<\/strong><\/p>\r\nFor students new to AI and data science, MATLAB provided a smooth on-ramp. Its intuitive syntax, built-in visualization, and integrated deep learning modeling\u00a0allowed the team to:\r\n<\/h6>\r\n\r\n \t- Experiment quickly with model architectures.<\/li>\r\n \t
- Easily visualize and analyze results.<\/li>\r\n \t
- Focus on their core engineering goals without losing time with figuring out the right tools.<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Looking Ahead<\/strong><\/p>\r\nThe team's work demonstrated the effectiveness of hybrid AI \/ physics-based approaches<\/a> in accelerating engineering simulations. Their strategy is not only technically elegant, but also directly applicable to real-word industrial settings where simulation time is a bottleneck.\r\n<\/h6>\r\nFuture improvements could include:\r\n<\/h6>\r\n\r\n \t- Improving the existing neural network<\/li>\r\n \t
- Exploring other decomposition techniques<\/li>\r\n \t
- Exploring other neural network architectures<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Final Thoughts<\/strong><\/p>\r\nThe INSA HDF team's contribution to the SIA Student Challenge highlights how emerging engineers can blend machine learning and numerical methods to push the boundaries of simulation science.\r\n<\/h6>\r\nCongratulations to Pierre, Joao, and their mentors for this outstanding achievement!\r\n<\/h6>\r\nWant to learn more about model reduction and AI engineering? Check out resources on reduced-order modeling and dive into deep learning with MATLAB.\r\n<\/h6>\r\nResources:<\/strong>\r\n<\/h6>\r\n\r\n \t- Finite Element Analysis with MATLAB<\/a><\/li>\r\n \t
- Physics-Informed Neural Networks<\/a><\/li>\r\n \t
- Reduced Order Modeling with MATLAB<\/a><\/li>\r\n<\/ul>\r\n
<\/h6>","protected":false},"excerpt":{"rendered":"<\/div>\r\nHighlights from the SIA Student Challenge - INSA Hauts-de-France Team\r\n\r\n\r\nThis blog post is from Giovanni Donati, Senior Consultant at MathWorks. \r\n\r\n \r\n\r\nAt the 2025 edition of the SIA... read more >><\/a><\/p>","protected":false},"author":194,"featured_media":18071,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[9],"tags":[],"_links":{"self":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056"}],"collection":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/users\/194"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/comments?post=18056"}],"version-history":[{"count":46,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions"}],"predecessor-version":[{"id":18236,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions\/18236"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media\/18071"}],"wp:attachment":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media?parent=18056"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/categories?post=18056"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/tags?post=18056"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
<\/h6>\r\nAs sponsors of the student challenge, MathWorks offered two tickets for the 24 Hours of Le Mans to the winning students!\r\n<\/h6>\r\n\r\n<\/h6>\r\nJoao and Pierre wearing their MathWorks t-shirts at the 24 Hours of Le Mans!<\/em>\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Smart Strategies for a Complex Problem<\/strong><\/p>\r\nTo solve this problem, the INSA HDF team developed and compared two alternative approaches:\r\n\r\n \t- Full neural network approximation<\/li>\r\n \t
- Model reduction supported by meta-modeling<\/li>\r\n<\/ul>\r\n\r\n
<\/h6>\r\nFigure 2:<\/strong> First neural network architecture, used for the initial approach<\/em>\r\n<\/h6>\r\nThe first approach led the INSA team to a good accuracy score (85.2%) but slow simulation time (about 1 hour for training and prediction). Using this model as a baseline for comparison, they improved the model and ultimately obtained impressive results, leading them to win the challenge! Let's dig into the winning approach.\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Model Reduction Supported by Meta-Modeling<\/strong><\/p>\r\nTheir second and more advanced solution, integrated a multiple-stage model reduction:\r\n<\/h6>\r\n\r\n \t- They applied Dynamic Model Decomposition (DMD) on each simulation, keeping only 15 key modes.<\/li>\r\n \t
- They used Singular Value Decomposition (SVD) to create a global reduced basis (300 modes).\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 3:<\/strong> Mode energies of SVD components<\/em><\/li>\r\n \t- They approximated the reduced coefficients by using a shallow neural network.\r\n
<\/h6>\r\nFigure 4:<\/strong> Neural network architecture of the winning solution<\/em><\/li>\r\n \t- They reconstructed the physical displacements.<\/li>\r\n<\/ol>\r\nThe full workflow is represented using the following principle scheme:\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 5:<\/strong> Principle scheme of the training and prediction process<\/em>\r\n<\/h6>\r\nBy adding extra data preprocessing steps, their new strategy achieved a slightly better accuracy and an impressive simulation speed-up compared to their first approach. They achieved:\r\n<\/h6>\r\n\r\n \t- Accuracy score: 86.2%<\/strong><\/li>\r\n \t
- Model Training and Prediction time: 150 seconds<\/strong><\/li>\r\n<\/ul>\r\nThe speed-up is 168x<\/strong>\u00a0compared to the first full machine learning trial!\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 6:<\/strong> Score and Pareto Front visualization provided by Renault for evaluating the quality of the models<\/em>\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Why MATLAB?<\/strong><\/p>\r\nFor students new to AI and data science, MATLAB provided a smooth on-ramp. Its intuitive syntax, built-in visualization, and integrated deep learning modeling\u00a0allowed the team to:\r\n<\/h6>\r\n\r\n \t- Experiment quickly with model architectures.<\/li>\r\n \t
- Easily visualize and analyze results.<\/li>\r\n \t
- Focus on their core engineering goals without losing time with figuring out the right tools.<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Looking Ahead<\/strong><\/p>\r\nThe team's work demonstrated the effectiveness of hybrid AI \/ physics-based approaches<\/a> in accelerating engineering simulations. Their strategy is not only technically elegant, but also directly applicable to real-word industrial settings where simulation time is a bottleneck.\r\n<\/h6>\r\nFuture improvements could include:\r\n<\/h6>\r\n\r\n \t- Improving the existing neural network<\/li>\r\n \t
- Exploring other decomposition techniques<\/li>\r\n \t
- Exploring other neural network architectures<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Final Thoughts<\/strong><\/p>\r\nThe INSA HDF team's contribution to the SIA Student Challenge highlights how emerging engineers can blend machine learning and numerical methods to push the boundaries of simulation science.\r\n<\/h6>\r\nCongratulations to Pierre, Joao, and their mentors for this outstanding achievement!\r\n<\/h6>\r\nWant to learn more about model reduction and AI engineering? Check out resources on reduced-order modeling and dive into deep learning with MATLAB.\r\n<\/h6>\r\nResources:<\/strong>\r\n<\/h6>\r\n\r\n \t- Finite Element Analysis with MATLAB<\/a><\/li>\r\n \t
- Physics-Informed Neural Networks<\/a><\/li>\r\n \t
- Reduced Order Modeling with MATLAB<\/a><\/li>\r\n<\/ul>\r\n
<\/h6>","protected":false},"excerpt":{"rendered":"<\/div>\r\nHighlights from the SIA Student Challenge - INSA Hauts-de-France Team\r\n\r\n\r\nThis blog post is from Giovanni Donati, Senior Consultant at MathWorks. \r\n\r\n \r\n\r\nAt the 2025 edition of the SIA... read more >><\/a><\/p>","protected":false},"author":194,"featured_media":18071,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[9],"tags":[],"_links":{"self":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056"}],"collection":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/users\/194"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/comments?post=18056"}],"version-history":[{"count":46,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions"}],"predecessor-version":[{"id":18236,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions\/18236"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media\/18071"}],"wp:attachment":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media?parent=18056"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/categories?post=18056"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/tags?post=18056"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
\r\n<\/h6>\r\nJoao and Pierre wearing their MathWorks t-shirts at the 24 Hours of Le Mans!<\/em>\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Smart Strategies for a Complex Problem<\/strong><\/p>\r\nTo solve this problem, the INSA HDF team developed and compared two alternative approaches:\r\n\r\n \t- Full neural network approximation<\/li>\r\n \t
- Model reduction supported by meta-modeling<\/li>\r\n<\/ul>\r\n\r\n
<\/h6>\r\nFigure 2:<\/strong> First neural network architecture, used for the initial approach<\/em>\r\n<\/h6>\r\nThe first approach led the INSA team to a good accuracy score (85.2%) but slow simulation time (about 1 hour for training and prediction). Using this model as a baseline for comparison, they improved the model and ultimately obtained impressive results, leading them to win the challenge! Let's dig into the winning approach.\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Model Reduction Supported by Meta-Modeling<\/strong><\/p>\r\nTheir second and more advanced solution, integrated a multiple-stage model reduction:\r\n<\/h6>\r\n\r\n \t- They applied Dynamic Model Decomposition (DMD) on each simulation, keeping only 15 key modes.<\/li>\r\n \t
- They used Singular Value Decomposition (SVD) to create a global reduced basis (300 modes).\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 3:<\/strong> Mode energies of SVD components<\/em><\/li>\r\n \t- They approximated the reduced coefficients by using a shallow neural network.\r\n
<\/h6>\r\nFigure 4:<\/strong> Neural network architecture of the winning solution<\/em><\/li>\r\n \t- They reconstructed the physical displacements.<\/li>\r\n<\/ol>\r\nThe full workflow is represented using the following principle scheme:\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 5:<\/strong> Principle scheme of the training and prediction process<\/em>\r\n<\/h6>\r\nBy adding extra data preprocessing steps, their new strategy achieved a slightly better accuracy and an impressive simulation speed-up compared to their first approach. They achieved:\r\n<\/h6>\r\n\r\n \t- Accuracy score: 86.2%<\/strong><\/li>\r\n \t
- Model Training and Prediction time: 150 seconds<\/strong><\/li>\r\n<\/ul>\r\nThe speed-up is 168x<\/strong>\u00a0compared to the first full machine learning trial!\r\n
<\/h6>\r\n\r\n<\/h6>\r\nFigure 6:<\/strong> Score and Pareto Front visualization provided by Renault for evaluating the quality of the models<\/em>\r\n<\/h6>\r\n \r\n<\/h6>\r\n
Why MATLAB?<\/strong><\/p>\r\nFor students new to AI and data science, MATLAB provided a smooth on-ramp. Its intuitive syntax, built-in visualization, and integrated deep learning modeling\u00a0allowed the team to:\r\n<\/h6>\r\n\r\n \t- Experiment quickly with model architectures.<\/li>\r\n \t
- Easily visualize and analyze results.<\/li>\r\n \t
- Focus on their core engineering goals without losing time with figuring out the right tools.<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Looking Ahead<\/strong><\/p>\r\nThe team's work demonstrated the effectiveness of hybrid AI \/ physics-based approaches<\/a> in accelerating engineering simulations. Their strategy is not only technically elegant, but also directly applicable to real-word industrial settings where simulation time is a bottleneck.\r\n<\/h6>\r\nFuture improvements could include:\r\n<\/h6>\r\n\r\n \t- Improving the existing neural network<\/li>\r\n \t
- Exploring other decomposition techniques<\/li>\r\n \t
- Exploring other neural network architectures<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Final Thoughts<\/strong><\/p>\r\nThe INSA HDF team's contribution to the SIA Student Challenge highlights how emerging engineers can blend machine learning and numerical methods to push the boundaries of simulation science.\r\n<\/h6>\r\nCongratulations to Pierre, Joao, and their mentors for this outstanding achievement!\r\n<\/h6>\r\nWant to learn more about model reduction and AI engineering? Check out resources on reduced-order modeling and dive into deep learning with MATLAB.\r\n<\/h6>\r\nResources:<\/strong>\r\n<\/h6>\r\n\r\n \t- Finite Element Analysis with MATLAB<\/a><\/li>\r\n \t
- Physics-Informed Neural Networks<\/a><\/li>\r\n \t
- Reduced Order Modeling with MATLAB<\/a><\/li>\r\n<\/ul>\r\n
<\/h6>","protected":false},"excerpt":{"rendered":"<\/div>\r\nHighlights from the SIA Student Challenge - INSA Hauts-de-France Team\r\n\r\n\r\nThis blog post is from Giovanni Donati, Senior Consultant at MathWorks. \r\n\r\n \r\n\r\nAt the 2025 edition of the SIA... read more >><\/a><\/p>","protected":false},"author":194,"featured_media":18071,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[9],"tags":[],"_links":{"self":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056"}],"collection":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/users\/194"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/comments?post=18056"}],"version-history":[{"count":46,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions"}],"predecessor-version":[{"id":18236,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions\/18236"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media\/18071"}],"wp:attachment":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media?parent=18056"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/categories?post=18056"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/tags?post=18056"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
<\/h6>\r\n
- \r\n \t
- Full neural network approximation<\/li>\r\n \t
- Model reduction supported by meta-modeling<\/li>\r\n<\/ul>\r\n\r\n
<\/h6>\r\nFigure 2:<\/strong> First neural network architecture, used for the initial approach<\/em>\r\n
<\/h6>\r\nThe first approach led the INSA team to a good accuracy score (85.2%) but slow simulation time (about 1 hour for training and prediction). Using this model as a baseline for comparison, they improved the model and ultimately obtained impressive results, leading them to win the challenge! Let's dig into the winning approach.\r\n
<\/h6>\r\n \r\n
<\/h6>\r\n
Model Reduction Supported by Meta-Modeling<\/strong><\/p>\r\nTheir second and more advanced solution, integrated a multiple-stage model reduction:\r\n
<\/h6>\r\n
- \r\n \t
- They applied Dynamic Model Decomposition (DMD) on each simulation, keeping only 15 key modes.<\/li>\r\n \t
- They used Singular Value Decomposition (SVD) to create a global reduced basis (300 modes).\r\n
<\/h6>\r\n
\r\n<\/h6>\r\nFigure 3:<\/strong> Mode energies of SVD components<\/em><\/li>\r\n \t
- They approximated the reduced coefficients by using a shallow neural network.
\r\n<\/h6>\r\nFigure 4:<\/strong> Neural network architecture of the winning solution<\/em><\/li>\r\n \t
- They reconstructed the physical displacements.<\/li>\r\n<\/ol>\r\nThe full workflow is represented using the following principle scheme:\r\n
<\/h6>\r\n
\r\n<\/h6>\r\nFigure 5:<\/strong> Principle scheme of the training and prediction process<\/em>\r\n
<\/h6>\r\nBy adding extra data preprocessing steps, their new strategy achieved a slightly better accuracy and an impressive simulation speed-up compared to their first approach. They achieved:\r\n
<\/h6>\r\n
- \r\n \t
- Accuracy score: 86.2%<\/strong><\/li>\r\n \t
- Model Training and Prediction time: 150 seconds<\/strong><\/li>\r\n<\/ul>\r\nThe speed-up is 168x<\/strong>\u00a0compared to the first full machine learning trial!\r\n
<\/h6>\r\n
\r\n<\/h6>\r\nFigure 6:<\/strong> Score and Pareto Front visualization provided by Renault for evaluating the quality of the models<\/em>\r\n
<\/h6>\r\n \r\n
<\/h6>\r\n
Why MATLAB?<\/strong><\/p>\r\nFor students new to AI and data science, MATLAB provided a smooth on-ramp. Its intuitive syntax, built-in visualization, and integrated deep learning modeling\u00a0allowed the team to:\r\n
<\/h6>\r\n
- \r\n \t
- Experiment quickly with model architectures.<\/li>\r\n \t
- Easily visualize and analyze results.<\/li>\r\n \t
- Focus on their core engineering goals without losing time with figuring out the right tools.<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Looking Ahead<\/strong><\/p>\r\nThe team's work demonstrated the effectiveness of hybrid AI \/ physics-based approaches<\/a> in accelerating engineering simulations. Their strategy is not only technically elegant, but also directly applicable to real-word industrial settings where simulation time is a bottleneck.\r\n
<\/h6>\r\nFuture improvements could include:\r\n
<\/h6>\r\n
- \r\n \t
- Improving the existing neural network<\/li>\r\n \t
- Exploring other decomposition techniques<\/li>\r\n \t
- Exploring other neural network architectures<\/li>\r\n<\/ul>\r\n
<\/h6>\r\n
Final Thoughts<\/strong><\/p>\r\nThe INSA HDF team's contribution to the SIA Student Challenge highlights how emerging engineers can blend machine learning and numerical methods to push the boundaries of simulation science.\r\n
<\/h6>\r\nCongratulations to Pierre, Joao, and their mentors for this outstanding achievement!\r\n
<\/h6>\r\nWant to learn more about model reduction and AI engineering? Check out resources on reduced-order modeling and dive into deep learning with MATLAB.\r\n
<\/h6>\r\nResources:<\/strong>\r\n
<\/h6>\r\n
- \r\n \t
- Finite Element Analysis with MATLAB<\/a><\/li>\r\n \t
- Physics-Informed Neural Networks<\/a><\/li>\r\n \t
- Reduced Order Modeling with MATLAB<\/a><\/li>\r\n<\/ul>\r\n
<\/h6>","protected":false},"excerpt":{"rendered":"
<\/div>\r\nHighlights from the SIA Student Challenge - INSA Hauts-de-France Team\r\n\r\n\r\nThis blog post is from Giovanni Donati, Senior Consultant at MathWorks. \r\n\r\n \r\n\r\nAt the 2025 edition of the SIA... read more >><\/a><\/p>","protected":false},"author":194,"featured_media":18071,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[9],"tags":[],"_links":{"self":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056"}],"collection":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/users\/194"}],"replies":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/comments?post=18056"}],"version-history":[{"count":46,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions"}],"predecessor-version":[{"id":18236,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/posts\/18056\/revisions\/18236"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media\/18071"}],"wp:attachment":[{"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/media?parent=18056"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/categories?post=18056"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blogs.mathworks.com\/deep-learning\/wp-json\/wp\/v2\/tags?post=18056"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
- Physics-Informed Neural Networks<\/a><\/li>\r\n \t
- Finite Element Analysis with MATLAB<\/a><\/li>\r\n \t
- Model Training and Prediction time: 150 seconds<\/strong><\/li>\r\n<\/ul>\r\nThe speed-up is 168x<\/strong>\u00a0compared to the first full machine learning trial!\r\n
- They approximated the reduced coefficients by using a shallow neural network.