In the midst of the AI boom and endless technological integrations, one technology organizations seem to still treat distinctly is simulation [1]. While both AI and simulation have been at the forefront of technological growth and digitalization, especially in the automotive industry, organizations like OEMs and automotive suppliers have yet to adopt a full integration of AI simulation. This is by no means a lack of technological compatibility; it is merely a case of technological integration in its infancy.
In 2023, researchers at Gartner positioned AI simulation within the so-called “innovation trigger” phase of their famous Gartner Hype Cycle™ of AI [2]. This phase is characterized by technology breakthroughs, proof-of-concept stories, and media publicity before it hits the so-called “peak of inflated expectations.” Yet, the pace at which AI simulation has been growing recently has shown significant potential and commercial viability, notably as the engineering simulation space has transformed with the advent of cloud-native simulation. The gap between expectation and practice is still wide: in SimScale’s State of Engineering AI research, 93% of engineering leaders expect AI to raise productivity, but only 3% say they are achieving that level of impact today, even as AI use in engineering and simulation workflows doubled in under a year [5].
What is AI simulation?
AI simulation is the use of machine learning models alongside physics-based simulation (CFD, FEA, thermal) to predict engineering results in seconds, automate simulation setup, or generate the data that trains AI models. A Physics AI model works by collecting relevant simulation data and training neural networks to build a model that can predict the outcome of a design or operating scenario that was not in the original data.
In engineering, AI simulation takes two forms. AI-assisted simulation uses AI to speed up or automate the simulation workflow itself. Simulation-assisted AI uses simulation output as training data for AI models. Both are covered below.
The term is also used outside engineering for agent-based simulations of human behavior, game and chat “AI simulators,” and synthetic data generation for software testing. This article covers engineering simulation only.
The impact of AI on engineering
As AI continues to evolve, it is likely to have a profound impact on engineering processes and practices. Many of the boring and repetitive jobs could be replaced by AI workflows. This is because the real benefit of AI is the speed at which results can be generated. However, it is still essential to review, check, and improve the data that is produced. This is why AI is not a threat to engineers, yet.
SimScale CEO David Heiny put it this way when opening the Engineering AI agent to the SimScale community: “AI is not going to replace hardware engineering jobs; it’s a task tool that will free up engineers to do the work they trained for, and dreamt about, in the first place.”
Engineers are however going to have to adapt and adjust to the changing environment or risk falling behind their peers and their competitors. There will be an increased importance for continued learning and experimentation with AI-based tools to ensure that the workforce is up to date with the latest developments. Engineers should not be threatened by AI, as if it is employed carefully and correctly it has the potential to make engineers a lot more effective.
Davide Russo
Co-Founder and Chief Innovation Officer, SunCubes
“SimScale gives us the capability to run complex, multi-physics analysis on our UAV designs and energy transfer systems in the cloud, with AI that helps us set up and interpret simulations faster than ever. For a team our size, this means we’re not blocked by inadequate tools and can move at startup speed, which is critical to our success.”
The convergence of AI and simulation
Engineering simulation has long been a fundamental tool for refining and validating designs. Throughout the latter half of the 20th century and the early years of the 21st century, simulation software has accumulated extensive data and calculations vital in testing and enhancing various engineering solutions. This has enabled millions of iterations of design optimization, but it has had limitations.
Burdened with hardware, computational, and software limitations, CAE simulation has had convoluted bottlenecks such as the inability to scale compute, accuracy-to-speed trade-offs, long simulation cycle times, and, more critically today, long simulation lead times. With the advent of cloud computing, cloud-native simulation has allowed organizations to resolve these bottlenecks and, in many cases, eliminate them altogether.
With cloud-native simulation, compute scales with the job, as engineers and designers can run as many simulations as needed in parallel without being hindered by hardware. As a result, accuracy does not need to be compromised, and simulation lead time is cut from weeks to hours. Running simulations online has also allowed organizations to overcome the data siloes between teams, enabling them to collaborate on simulation projects in real time by simply sharing the projects at a click of a button, thus facilitating the company-wide deployment of simulation.
Today, AI is changing how we approach simulation again. The convergence of AI and simulation is changing simulation methodologies by infusing them with AI-driven insights and capabilities. The essence of AI simulation lies in its ability to speed up and improve the simulation process, making it more efficient and accessible to engineers and designers across the entire design cycle, not merely as a validation step at the end.
At the heart of AI simulation are two approaches: AI-assisted simulation and simulation-assisted AI.
AI-assisted simulation
Here, AI algorithms are deployed to tune simulation parameters and processes, shortening design iterations and speeding up solution convergence. This AI-driven optimization speeds up the simulation process and improves its accuracy and effectiveness.
Analysts at Gartner [3] recommend complementing simulation with AI by applying the following:
- Deep learning to accelerate simulation
- Generative AI to augment simulation
As a result, simulation projects would benefit from AI’s capabilities to increase the speed of executing complex simulations or fill gaps in CAE and simulation models using generative AI. Today, only a cloud-native simulation infrastructure can enable full advantage of AI, as all simulations are immediately ready for AI training on cloud GPUs. As new data is added, retraining of AI models is straightforward. On SimScale’s cloud-native platform, for example, Physics AI models are deployed next to CFD simulations, with AI results used for near-instantaneous evaluations and CFD for validation.
Simulation-assisted AI
Simulation-assisted AI uses simulation outputs to enrich AI models. By integrating insights from simulations, AI models can be trained to understand complex systems better and predict their behavior with greater precision. This symbiotic relationship between simulation and AI lets engineers and researchers draw on the strengths of both disciplines, ultimately leading to more reliable and better-informed engineering solutions.
Supporting AI with simulation data helps improve business decision-making and overcome the common challenge of real-world data scarcity by providing a data-rich simulated environment for synthetic data generation and reinforcement learning. It is only a matter of time before AI becomes a natural and complementary addition to the use of synthetic data in areas such as automotive, supply chain, electronics manufacturing, and industrial equipment manufacturing.
A step change in engineering design and simulation
The simulation industry has already been going through a state of rapid change and development with the introduction of truly cloud-native engineering simulation like SimScale. However, AI and ML models have the potential to further change and evolve how simulation is used in the product design cycle (McKinsey).
Traditionally, physics-based simulation has been used late in the design cycle to validate designs and potentially reduce the reliance on physical prototyping. However, in a report by McKinsey, they report from their surveys of R&D leaders in engineering that the business case for using engineering simulation is moving more towards faster-time-to-market and reduced product cost.
Consider an example of a project to design a water pump to a specific specification for a client:
- Two-month window for pump design.
- First month: Design and CAD geometry based on experience and company precedent.
- Second month: If the CFD team has capacity they simulate the design and determine its performance against the specification.
- There may be a couple of feedback loops required for CAD cleanup operations, etc.
- This leaves limited time in the final week or two for design improvements (one iteration) based on simulation results.
Proposed alternative with a Physics AI model:
- Designer uses existing data for AI prediction of pump performance.
- Early simulation-informed design changes are possible as soon as the designer has a concept CAD model available.
- This opens up the possibility for generative engineering workflows that automatically explore design spaces or use automated design approaches.
- Shorter design process duration, reducing project overrun risk and late delivery risk.
- Allows for final simulation validation of the design in a shorter or equivalent time scale.
This is no longer hypothetical. Convion, a solid oxide fuel cell and electrolyzer manufacturer, ran a large-scale design of experiments on SimScale with hundreds of CFD simulations in parallel, trained a Physics AI model on the results, and used it to find the best-performing configuration for a component handling high-temperature process gases. The model predicted performance within 5% of high-fidelity CFD, and the team had a new design in under an hour. “The design was complex, and using traditional simulation-driven optimization to find the best-performing configuration would have taken months,” said Armin Narimanzadeh, Manager, Thermofluids & Simulations at Convion. “We now have an AI model that can generate a new optimized design in under an hour, and I have complete confidence in the results.”
How Physics AI models are trained: from DoE to prediction in seconds
A Physics AI model is trained on simulation results and then used in place of the solver for design exploration. On SimScale the process has four steps.
- Generate data at scale. Run a design of experiments over the design parameters, with hundreds of variants solved in parallel in the cloud. Existing simulation projects can be used as training data as well. SimScale requires at least 20 completed simulation runs of a single analysis type to train a model; more runs give higher prediction accuracy (documentation).
- Train the model. The model learns the relationship between design parameters (geometry, boundary conditions, operating point) and the physical results (surface pressure and velocity fields, from which performance quantities are derived). Training runs on cloud GPUs inside the platform.
- Use AI and solver side by side. The AI model and the full solver share the same setup, so a design can be evaluated with the model in seconds and then re-run in full CFD or FEA for validation without rebuilding the case.
- Run inference anywhere. Predictions can be requested from the browser, from a connected CAD tool, or through the API in an automated optimization loop. Each model is versioned and shared within the organization, so the same model can be re-used the next time the design changes.
Armin Narimanzadeh
Manager, Thermofluids & Simulations, Convion
“Over time, as we train more and more Physics AI models, it would just make everything more fast, agile and more accessible.”
For teams without their own dataset, a pre-trained Physics AI foundation model for centrifugal pumps, built with NVIDIA PhysicsNeMo and trained on thousands of validated CFD simulations, gives performance predictions in seconds and can be fine-tuned on your own designs.
Train a Physics AI model on your own simulation data
Physics AI runs next to CFD, FEA, and thermal solvers on the same cloud-native platform.
AI surrogate models vs. physics solvers
An AI surrogate model and a physics solver answer the same question by different means, and each has a range where it is the right tool.
| Criterion | Physics solver (CFD, FEA, thermal) | Physics AI surrogate model |
|---|---|---|
| What it does | Solves the governing equations on a mesh for one case | Learns the mapping from design inputs to results from many solved cases |
| Time per evaluation | Minutes to hours | Milliseconds to seconds |
| Upfront cost | Setup and mesh per case | A training dataset (a DoE) plus training time, once |
| Accuracy | Solver fidelity, limited by mesh and models | Close to the solver inside the trained design space (Convion: within 5% of CFD) |
| Valid range | Any geometry and physics the solver supports | Only the design space covered by the training data |
| Best use | Final validation, unusual cases, new physics | Design exploration, optimization loops, RFQ response, real-time prediction |
The practical pattern is hybrid: explore thousands of variants with the surrogate, then validate the shortlisted designs with the solver. Convion validated its approach this way, running a full CFD-driven optimization alongside the Physics AI model and arriving at an identical result.
Simulation-driven design with Physics AI and cloud-native simulation
The convergence of AI and cloud-native simulation has changed how simulation-driven design is done. Physics AI, driven by machine learning models trained on previous relevant data, changes how and when engineering simulation is used in the design cycle.
For instance, Graph Neural Networks (GNNs) stand out as a particularly powerful method in this domain. They use structured node-based data akin to the mesh used in engineering simulations like FEA or CFD. However, the hallmark of GNNs lies in their ability to deliver simulation results within seconds while maintaining reasonable accuracy.
Physics AI comes with its own set of challenges. The quality of outcomes heavily relies on the quality of input data used for training. For many organizations, data collection and sorting pose significant hurdles in getting the full value from Physics AI models. This is where cloud-native simulation can be the perfect infrastructure for Physics AI.
With Physics AI built into a cloud-native platform, data generation and preparation happen in one place, with data that is 100% compatible and fully traceable. The training of AI models runs on existing GPU infrastructure via high-performance computing (HPC) and is fully accessible and controllable. All this enables a flexible and interchangeable use of AI and PDE solvers in a single interface. That is exactly what SimScale is offering today.
SimScale’s first AI-based physics predictions reached the platform in 2023 through a partnership with NAVASTO (see the release webinar). Since 2025, Physics AI model training and inference are native to the platform and built on NVIDIA PhysicsNeMo, so engineers can train on their own simulation data or start from a foundation model without managing data pipelines or GPU infrastructure themselves.
This combination of AI and cloud-native simulation infrastructure speeds up the design iteration process and gives every engineer access to simulation methods that used to be specialist-only. As more organizations adopt it, they stand to gain efficiency and accuracy in their design workflows.
4 advantages of AI simulation
AI simulation offers a multitude of benefits for engineering design, especially regarding process efficiency, accuracy of simulation results, and decision-making. Here are four key advantages that AI simulation provides:
Faster innovation: One of the most significant benefits is the speed at which AI-assisted simulations operate. Unlike traditional methods, AI can swiftly analyze vast amounts of data from past simulations and provide results almost instantly. By identifying and examining complex patterns within the data, AI can shorten the design and optimization process.
Simulation for every engineer: AI-assisted simulations can also help make simulation accessible to every engineer. Previously reserved for simulation experts, AI simulation can build on top of what cloud-native simulation offers and help make simulation accessible to non-experts. This accessibility supports collaboration by enabling a broader range of design engineers to engage in design analysis, allowing for early simulation usage in the design cycle.
Comprehensive data: AI simulations integrate multiple models to offer a more comprehensive understanding of complex systems. These simulations provide thorough representations, allowing engineers to gain deeper insights into the behavior of the systems they are designing. This comprehensive approach improves decision-making and facilitates the creation of more reliable and efficient designs.
Constant learning: This is an inherent benefit of AI, with its iterative nature promoting continuous improvement. Engineers can refine and improve their designs with gradually fewer limitations, which leads to better solutions and outcomes, ensuring that designs meet evolving requirements and standards.
How to validate AI simulation results
A Physics AI prediction is only as reliable as its training data and the range it was trained on. Four checks keep AI results usable in a design review.
- Hold out validation cases. Keep a subset of solved designs out of training and compare the model’s predictions against them. Report the error on the quantities you design against (pressure drop, peak temperature, stress), rather than only an averaged field error.
- Stay inside the trained design space. A surrogate interpolates well and extrapolates badly. If a candidate design sits outside the parameter ranges of the DoE, treat the prediction as a hint and run the solver.
- Re-run the finalists with the solver. Use the model to reduce thousands of candidates to a handful, then run full CFD or FEA on those before committing the design.
- Track data quality. Training data inherits the mesh, turbulence model, and boundary condition choices of the original simulations. Inconsistent setups across the training set lower model accuracy, which is one reason a single, shared simulation platform with one setup per analysis type is the practical prerequisite for AI simulation.
Can AI simulation replace physical testing?
No. AI simulation reduces the number of design iterations that reach a prototype, in the same way that simulation has done for decades, but material variability, manufacturing tolerances, certification requirements, and failure modes not represented in the training data still require physical tests. What changes is the sequence: with predictions available in seconds, more of the exploration happens before the first prototype, and the test program validates a design that has already been narrowed down.
Frequently asked questions
AI simulation is the use of machine learning models together with physics-based simulation such as CFD and FEA. AI-assisted simulation speeds up or automates the simulation workflow; simulation-assisted AI uses simulation results as training data for AI models that then predict results in seconds.
A physics solver computes the result for one case by solving the governing equations on a mesh, which takes minutes to hours. An AI surrogate model is trained on many solved cases and predicts results for a new design in seconds, but only within the design space it was trained on. In practice the two are used together: the surrogate for exploration, the solver for validation.
On SimScale, a model can be trained from at least 20 completed simulation runs of a single analysis type. Accuracy improves with more runs, so most teams generate training data with a design of experiments of hundreds of variants run in parallel. Pre-trained foundation models, such as the centrifugal pump model, remove the need to build a dataset first.
No. AI models are trained on CFD and FEA results and depend on them for both training data and validation. They replace solver runs for repetitive design exploration inside a known design space, and the solver remains the reference for final validation and for cases outside that space.
Accuracy depends on the training data and the design space. In Convion’s published case, a Physics AI model trained on a large-scale CFD design of experiments predicted performance within 5% of high-fidelity CFD, and a parallel CFD-driven optimization arrived at the same design. Results for other applications should be validated the same way before use.
Set up your own cloud-native simulation and Physics AI training in the browser by creating an account on SimScale. No installation, special hardware, or credit card is required.
References
- Gartner (2023). Innovation Insight: AI Simulation. https://www.gartner.com/en/documents/4037399
- Gartner (2023). What’s New in Artificial Intelligence from the 2023 Gartner Hype Cycle. https://www.gartner.com/en/articles/what-s-new-in-artificial-intelligence-from-the-2023-gartner-hype-cycle
- Cosmotech (2023). Gartner Highlights Powerful Combination Of AI & Simulation. https://cosmotech.com/resources/article/gartner-highlights-powerful-combination-of-ai-simulation/
- von Rueden, L. et al. (2020). Combining Machine Learning and Simulation to a Hybrid Modelling Approach: Current and Future Directions. International Symposium on Intelligent Data Analysis (IDA). 548-560. DOI:10.1007/978-3-030-44584-3_43
- SimScale (2025, 2026). State of Engineering AI. https://www.simscale.com/research-reports/state-of-engineering-ai-2026/ and https://www.simscale.com/blog/the-engineering-ai-ambition-execution-gap-what-our-new-global-survey-reveals/
- McKinsey & Company. Unveiling the next frontier of engineering simulation. https://www.mckinsey.com/capabilities/operations/our-insights/unveiling-the-next-frontier-of-engineering-simulation
