Introduction
Artificial intelligence (AI) is now part of many technical disciplines, and simulation is no exception. The central question we explored in a recent podcast with Dr. Çağlar Aytekin, Lead AI Developer at Quanscient, was “What is the role of AI in multiphysics simulation today, and how might it develop in the near future?”
The discussion covered several dimensions, from physics-informed neural networks (PINNs) to surrogate models and hybrid approaches, as well as engineering-side applications such as anomaly detection and documentation assistance.
The role of AI in simulation workflows
AI is enhancing simulation in two broad ways:
-
Computation acceleration: Neural surrogate models can approximate the results of classical solvers with significant speed improvements. They allow engineers to explore design spaces faster and test more variations in less time.
-
Automation of engineering tasks: AI tools can assist with geometry setup, assigning materials, checking models before simulation, and detecting anomalies. This reduces manual effort and prevents wasted computational resources.
According to Dr. Aytekin, both areas are developing in parallel and should be considered complementary rather than competing.
Engineering-side AI
On the engineering side, AI applications often involve generative models such as large language models (LLMs). Their role is to support human engineers by:
- Helping with problem formulation
- Assisting with documentation
- Double-checking user setup, such as via performing anomaly detection
- Interpreting simulation outputs and results
In more advanced use cases, AI agents can carry out complex tasks before and after simulations. Such tasks may include both optimizing and performing geometry forming, meshing, simulation and post-processing, via language commands from the user.
Simulation-side AI
The simulation side is where AI interacts most directly with the physics of the problem. Dr. Aytekin distinguishes three main approaches:
-
Physics-Informed Neural Networks (PINNs): These models embed physical equations and boundary conditions into the neural network’s training process. Instead of approximating outputs from a dataset, PINNs aim to solve the governing equations directly.
While they are still in the research phase and not yet as accurate as finite element methods (FEM), they offer a different paradigm that could eventually provide comparable accuracy with faster runtimes.
-
Neural surrogates and neural operators: Surrogate models learn input-output mappings from data generated by classical solvers. Once trained, they can provide results much more quickly than FEM or PINNs.
The main challenge is the need for very large, high-quality datasets, which are often unavailable.
-
Hybrid approaches: Hybrid workflows combine surrogates with traditional solvers and/or PINNs. Training surrogate models in a physics-informed manner drastically decreases the data need of these models. Once trained, surrogates are used to explore many candidate designs rapidly, and the most promising candidates are validated with FEM or other numerical methods. This balances speed and reliability.
Challenges and limitations
Each method faces its own set of challenges:
-
PINNs require expensive GPU training per problem and are sensitive to the choice of optimizer. Traditional optimizers like Adam may not be sufficient; physics-specific methods such as Hessian-based optimizers may be necessary. Accuracy remains a limitation compared to established solvers.
-
Neural surrogates are limited by the availability of training data. Without sufficient diversity and quality, their predictions may be unreliable outside the training distribution.
-
Hybrid methods mitigate some issues but depend on surrogates producing at least some promising candidates for validation.
Quanscient Allsolve's approach
Quanscient Allsolve is focusing on both solver-side and engineering-side AI. The company is developing:
-
A simulation agent that supports documentation, scripting, anomaly detection, and related tasks.
-
Neural surrogates, with early experiments already showing promising results.
A key advantage is the ability to generate training data at scale using Quanscient Allsolve's own simulation infrastructure. This removes reliance on limited public datasets and allows for the training of more capable models.
The near and medium-term look
Looking three to five years ahead, Dr. Aytekin expects:
- Broader use of hybrid simulation methods.
- More widespread verification tools to check AI-generated results, either through interpretability methods or by comparison with classical solvers.
- Greater adoption of AI assistants for routine engineering tasks, from geometry definition to post-processing of results.
Beyond that, in the five- to ten-year horizon, foundational physics models could become important. Similar to how LLMs serve as general-purpose text models, foundational models for physics may serve as starting points for a wide range of simulation tasks.
Personal perspective
When asked what excites him most, Dr. Aytekin points to physics-informed neural networks. He emphasizes their potential to eventually replace methods that have been in use for half a century with approaches grounded in deep learning.
Although PINNs remain immature, he believes that with the right neural architectures and training strategies, they could have a major impact.
Conclusion
AI in multiphysics simulation is progressing on two fronts: solver-side applications such as PINNs and surrogates, and engineering-side tools that automate workflows and reduce errors. Each approach has distinct advantages and limitations, but together they are shaping a new landscape for simulation.
As Dr. Aytekin highlights, the next decade will likely bring incremental adoption of hybrid methods, more effective verification of AI outputs, and eventually, the development of foundational models for physics. Quanscient Allsolve’s ability to generate its own large-scale training datasets positions it to play a significant role in this transformation.
