News
18.06.26
AI in Simulation: How Physics AI is revolutionising CAE processes
Finite element analyses have been the tool of choice for design engineers for decades. They are used to model structures, heat transfer and fluid flow, and to simulate and understand how designs behave under loads. It works. But it is time-consuming. A complex model can take hours or even days to compute. Parametric studies – for example, determining the optimal wall thickness – require dozens of simulations. The time to insight is long.
Artificial intelligence is changing this dynamic. In particular, approaches such as physicsAI combine traditional CAE with machine learning. The result: simulations are 100 times faster. Parametric studies become real-time analyses. Design engineers can explore designs interactively – rather than waiting.
This is not just hype. It is a genuine transformation of CAE processes. This article explains how – and where physicsAI is already making an impact today.
Why AI is now becoming relevant in CAE
CAE is now an established discipline. Every major design consultancy has FEM tools such as Simcenter or ANSYS. However, two problems persist:
Problem 1: Computation time. A detailed structural simulation with millions of elements takes 8–48 hours. Simulating heat transfer in a complex component can take days. Turbulence flow simulations are even more time-consuming. If a design engineer changes a parameter in the morning, they receive the result in the afternoon – too late for interactive decision-making.
Problem 2: Throughput. A parametric study requires 100 variants. At 20 hours per simulation, that equates to 2,000 hours of computation time. This means weeks of waiting for answers. Design engineers rarely carry this out; instead, they make decisions based on intuition or simple rules of thumb.
Consequently, CAE is often limited to validation: design creates a concept, CAE validates it. If problems arise, the process is iterated. This is a feedback loop, but it is neither tight nor fast.
AI-based simulation changes this approach. If you can perform a parameter scan in seconds, the workflow changes fundamentally. Design engineers can work exploratively: ‘What happens if I increase the wall thickness from 2 to 2.5 mm?’ Seconds later, they know the answer. This opens up design spaces that were previously unexplored.
This is strategically important: companies with fast simulation iterate more. They produce better designs. They bring products to market faster. This is not a niche advantage – it is a fundamental competitive advantage.
The technical reason why AI works now: deep learning has proven to be powerful for physical systems. With sufficient training data, a machine learning model can learn how a system behaves physically – faster and with sufficient accuracy.
How physicsAI works
PhysicsAI is not magic. It is a structured approach: classical simulation + machine learning + validation.
Step 1: Classical simulation. You define a CAE model – geometry, material, boundary conditions, mesh. You carry out a series of classical simulations with varying parameters. This can involve between 100 and 1,000 runs. This is time-consuming, but a one-off investment.
Step 2: Data collection and pre-processing. The results of all simulations are collected: input parameters (wall thickness, material, load, etc.) and output (stress, displacement, temperature, etc.). This data is cleaned and normalised.
Step 3: Machine learning training. A neural network is trained on the data. The network learns the relationship: if I input these parameters, these are the results. Training takes anywhere from minutes to hours, depending on the model’s complexity.
Step 4: Validation. The trained model is validated against traditional simulations. Typical accuracy ranges from 95–99 per cent, depending on the type of application. If accuracy is too low, you will need more training data or a better model.
Step 5: Deployment. The trained model is exported and integrated into CAE workflows or made available as a standalone app.
The advantage is clear: following a one-off investment, you can evaluate a trained model 10,000 times faster than with traditional simulation. A FEM run that normally takes 20 hours is reduced to 0.7 seconds.
An important detail: the ML model is not a black box. PhysicsAI approaches often use ‘physics-informed neural networks’ – the neural network is trained using the laws of physics, not blindly with data. This makes the results more interpretable and robust.
Comparison of traditional simulation and AI models
When is which approach appropriate?
Traditional FEM remains important for: One-off complex analyses involving new geometries. Validations requiring the highest accuracy. Problems where physical non-linearity or complexity is difficult to predict. Certification processes that require traditional methods.
AI models are better suited to: Parametric studies – many variations of a design. Real-time feedback in design tools. Optimisation – when you need to test thousands of variants to find the best one. Rapid preliminary investigations prior to detailed analyses. Interactive design exploration.
In practice, both are combined: traditional FEM generates the training data. The AI model is trained. In the daily workflow, the design engineer uses the AI model for quick checks. If a design shows promise, it is validated in detail using traditional FEM.
A concrete example: you are designing a cross-member with an unknown optimal cross-section. Conventional approach: you carry out FEM analyses on five cross-section variants. This takes 100 hours. You select the best one. With AI: You train a model based on other similar beams (or quickly generated using traditional FEM). You test 500 variants in seconds. You select the best one. You carry out detailed validation. Result: A better design, completed more quickly.
Accuracy is a key consideration. ML models have errors, typically 1–5 per cent. This is acceptable for design exploration. However, for final approval, you’ll want traditional FEM validation. The combination gives you the best of both worlds: speed and reliability.
Practical applications today
PhysicsAI is not just a thing of the future. It is already being used in companies today. Here are a few real-life examples:
Example 1 – Structural optimisation: An automotive supplier needs to make a beam lighter without losing stiffness. A traditional optimisation process would take weeks – involving hundreds of FEM runs. With PhysicsAI: A trained model shows within seconds how changes to the cross-section affect stress and stiffness. The engineer explores options iteratively. A new, optimised design emerges in days, not weeks. This not only saves time but often results in better designs, as more variants have been tested.
Example 2 – Thermal analyses: A manufacturer of electronic heat sinks needs to be able to quickly demonstrate how a custom heat sink reduces a chip’s temperature. This is a common customer enquiry. Traditional thermal FEM takes hours. With an AI model: the customer enters the chip specifications and heat sink dimensions, and the model returns the temperature instantly. The sales team can provide a quote over the phone. This is not only faster, but it also represents a new service.
Example 3 – Flow optimisation: A pump manufacturer is testing new impeller geometries. Each traditional CFD simulation takes 30 hours. A physicsAI model takes seconds. Design engineers can test 100 geometries per day instead of two per week. The overall pace of development is multiplied.
Example 4 – Material selection and cost optimisation: A plastic component designer is testing different material grades and wall thicknesses against requirements (stiffness, strength, temperature resistance). With physicsAI, the best combination can be found quickly and cost-effectively.
These examples show that the applications are real and address genuine time-to-market challenges.
Borders & the Future
PhysicsAI is powerful, but not limitless. It is important to be aware of its limitations:
Limitation 1: Extrapolation. An AI model learns from training data. If you feed it parameters that lie outside the training range, accuracy can drop significantly. Example: You train the model on wall thicknesses of 1–5 mm. You test it with 10 mm. The model can extrapolate, but poorly. Solution: Careful training design and validation against classical FEM at the boundaries.
Limitation 2: New geometries. A model trained on cylindrical beams cannot automatically be applied to rectangular ones. You need new training. This requires effort. However, in the medium term, parametric models are being developed that can generalise across geometries.
Limitation 3: Coupled physics. A model trained for structural stress cannot automatically handle temperature coupling as well. Multi-physics systems are more complex. They require more extensive training.
Limitation 4: Unexpected effects. An AI model learns from what is in the training data. A rare phenomenon that did not occur during training will not be predicted. This is a classic machine learning problem. Validation and sanity checks are necessary.
Future: Several developments will make AI simulation even better. Firstly: generalised models. Instead of training one model per geometry type, universal models may exist in future – ‘meta-models’ that are applicable to many geometries. Secondly: federated learning. Instead of every company training its own model, large firms will be able to use shared data pools to produce better models. Thirdly: integration with design tools. CAD tools such as NX will integrate AI simulation natively – designers will be able to change parameters and see the AI simulation results immediately. This is a revolution in workflow.
The next decade will see CAE transform from a validation tool into a design platform. This will be made possible by AI and simulation.
Getting started in your business: Practical steps
How do you get started with physicsAI?
Step 1: Pilot project. Select a specific CAE problem with a high volume of repetitions. For example: structural optimisation of a component family, or thermal analyses of standard designs. This is a project with measurable benefits.
Step 2: Collect data. Run traditional simulations for the pilot – 100 to 500 variants. This is time-consuming, but a one-off task. You’ll generate training data.
Step 3: Train the model. A data science partner will help to train and validate a machine learning model. This takes weeks, not months.
Step 4: Integration. The trained model is integrated into your CAE workflow. Possibly as a plug-in in Simcenter, or as a standalone Python script that engineers can run.
Step 5: Measure ROI. After three months, measure: How much time did the AI save? How many more variants was your team able to test? Were better designs developed? This will indicate whether to scale up.
Organisational: Design engineers must learn to use the tool. That is training. It is also a cultural issue – building trust in an AI model takes time. Transparency and validation help here.
Costs: A complete AI simulation solution for a specific problem costs €50K–200K in development and tooling. The ROI is usually positive within 6–12 months because development times are reduced.
Conclusion: CAE transformation through AI
The traditional CAE method – building an FEM model, running the simulation, waiting for hours, interpreting the results – is not going to disappear. But it will be supplemented and, in some cases, replaced by AI models that take mere seconds.
This is not a gradual progression. It is a shift in how engineers design. With slow CAE, parametric studies are prohibitively expensive. With fast CAE, they are standard practice. With slow CAE, engineers wait for feedback. With fast CAE, they receive feedback immediately.
The economic reality: companies that adopt AI simulation early develop products faster, more cheaply and to a higher standard. This translates into market advantages: faster time-to-market, a higher rate of innovation and a better cost position.
This is no longer the exception for Google-scale labs. It is practical and cost-effective for SMEs and large corporations in mechanical engineering, the automotive sector and medical technology.
d.u.h.Group supports you through this transformation. We integrate Simcenter with AI platforms. We help you collect training data and train models. We set up workflows that make AI simulation the standard in your design process.
Get in touch with us. We’ll show you which of your CAE processes can be transformed using physicsAI – and what specific benefits await.
Topics
Technologies
Mehr entdecken