November 19, 2020
Current modeling and simulation software use either linear or nonlinear solvers that employ the traditional numerical methods approach. The solvers decompose the model into finite elements, approximate the solution through polynomial curve fitting, and translate the model into a system of equations that are solved using adequate methods for sparse matrix operations.
Since the introduction of neural networks, artificial intelligence (AI) has steadily improved the job performance of Simulation Analysts. Regardless of one’s approval of AI’s presence, it is undeniable that AI is becoming more ubiquitous and versatile. The adoption of AI, which is accelerated by cloud computing, is considered the future of efficient computer modeling and simulation.
Some of the most widely used implementations of AI include:
- Infectious disease pandemics
- Pattern recognition
The question “Can AI help me be more effective and improve my performance?” is becoming common among Simulation Analysts. This question is considered relevant in the context of standardization, cloud computing, and predictive modeling.
Evidence suggests that some of the most tedious sub-steps in the simulation workflow are already incorporating a high degree of automation, a precursor for AI entering this field.
From an FE Analyst’s point of view AI taking over for standard simulation solvers can be broken down into the three main simulation processes discussed below.
Pre-processing, the activity needed to prepare the model/simulation for analysis by cleaning up CADs, assigning constitutive models, and loads, is ripe for AI takeover.
From auto mesh generators to statistical modeling involving complex algorithms, AI starts to establish itself as the natural progression in simulation and modeling. While the progress seems slow today, there is evidence that software developers are working on AI-enabled versions of their codes that can handle simple models.
Solvers, the core of the simulation process that takes the pre-processed models and translates them into a mathematical system of equations, is currently an active target of AI and neural networking.
In the vast realm of solvers, current computer simulation programs invoke specific numerical methods that best fit the resulting model setup and automatically adapt if the solutions seem to diverge towards unnatural trends. AI-enhanced solvers rely heavily on data analytics to accelerate equation-based modeling.
Solutions such as evolutionary polynomial regression (EPR) that use data-driven methods based on evolutionary computing to search for polynomial structures representing a system are already integrated into some FE analysis solvers. The process of training the EPR, computing the stiffness matrix using the trained EPR model, and incorporating the EPR constitutive model in a commercial finite element code has begun. This approach promises to reduce simulation cost (data analytics still requires considerable computation effort) and deliver a scalable and robust algorithm that can be implemented in commercial FE software.
Post-processing, the realm once viewed as the ultimate frontier for automation and AI, is becoming more and more palatable for pattern-recognition and refinement, tasks that are AI’s forte.
Converging to a mesh-independent solution requires a detailed analysis of the results, which usually involves extensive data analysis, perfect breeding grounds for neural networks and AI. Automation can eventually lead to AI integration in the results interpretation process that is currently a team-based approach in most commercial organizations, helping with rapid prototyping and, soon, additive manufacturing.
The figure below shows an example of an AI-enhanced solver map:
It is difficult to unequivocally provide a stern prediction for the role of AI in the future of simulation solvers. Will it take over? Will it augment our strengths? Will it fizzle out before making its mark on modeling and simulation?
The question shall remain unanswered today, but it is not far-fetched to imagine that an expert system can solve a simple FE problem as AI evolves. In the process, AI can learn to model complex systems and guide the analyst through the steps freeing up time to interpret results, help with quick prototyping, and creatively implement the solutions into final product design.