TL;DR
This paper explores the use of Genetic Programming (GP) to automatically discover constitutive models for materials, a crucial step in simulating structural behavior. By treating model discovery as a search problem, GP successfully identified both one-dimensional rheological models and three-dimensional hyperelastic strain energy functions, demonstrating its potential for accelerating materials science and engineering design.
What Problem Does This Paper Address?
Accurately predicting how structures will behave under load is fundamental to engineering design. This requires detailed knowledge of the material's constitutive properties – its "constitutive law" – which describes its mechanical behavior. For well-understood materials, engineers can often adjust existing models. However, for new or complex materials (like advanced polymers or composites), developing accurate constitutive laws is a significant challenge. Traditional methods involve time-consuming trial-and-error, microscopic analysis, or adjusting existing models, all of which can be complex, computationally expensive, or insufficient for highly nonlinear behaviors. The paper addresses the need for an automated approach to identify these constitutive laws directly from experimental data.
Methods
The paper applies Genetic Programming (GP) to two distinct problems in macro-mechanical modeling:
- One-Dimensional Rheological Models: These models describe a material's dynamic behavior (viscoelasticity, plasticity) as an assembly of basic components (springs, sliders, dashpots) connected in series or parallel. GP is used to evolve tree structures representing these assemblies. An interpreter then translates these trees into equivalent systems of differential equations. Key aspects include:
- Search Space: Defined by a set of functions (serial/parallel connectors) and terminals (spring stiffness, slider stress threshold, dashpot viscosity), with a "language bias" to reduce representation degeneracy.
- Solver: A numerical solver that computes the material's response to applied strain histories by solving a system of linear equations derived from the rheological model.
- Fitness Function: Measures the difference between the model's predicted stress and the experimentally observed stress.
- Evolutionary Operators: Standard crossover and mutation, augmented with a "surface mutation" (Gaussian noise addition) and a random hill-climber to optimize numerical constants efficiently.
- Iterative Fitness Refinement: To manage computational cost, the fitness evaluation is gradually refined, starting with fewer experiments or time steps and increasing precision over generations.
- Three-Dimensional Hyperelastic Strain Energy Functions: For hyperelastic materials (like rubber), behavior is described by a strain energy function, $W$, which depends on invariants of the deformation gradient. GP is used to discover this function.
- Search Space: Standard GP function set ({+, -, *}) and terminals (material invariants, random constants).
- Numerical Model: A finite element method (FEM) solver is used to compute the structure's displacement response for a given strain energy function and loading case. This involves solving a system of PDEs.
- Inverse Problem: The goal is to find the $W$ that minimizes the mean square error between experimental displacements and those predicted by the FEM for a given $W$.
- Constraints: To ensure physically plausible models, GP incorporates constraints:
- A posteriori checks for limits of $W$ as variables go to infinity.
- A no-loading condition derived from material theory.
- Derivatives: A key aspect is the use of derivatives of the strain energy function within the FEM analysis, which GP is well-suited to handle.
Datasets
The paper uses simulated experimental data for both types of models.
- One-Dimensional Rheological Models: Simulated data based on a known rheological model of polyethylene was used. The "experimental" data was generated by a PDE solver for this known model.
- Three-Dimensional Hyperelastic Strain Energy Functions: Simulated data was generated using a simplified form of Ogden materials (Mooney-Rivlin materials) with known coefficients. This data was then used to train the GP to identify the strain energy function.
Key Insights and Ideas
- GP as a General-Purpose Model Discovery Tool: The paper strongly advocates for GP as a powerful, non-parametric tool for discovering complex material models where traditional methods fall short. It can discover both the structure and parameters of the model simultaneously.
- Handling Numerical Constants: Optimizing floating-point constants within GP is a significant challenge. The paper introduces and evaluates specialized operators like "surface mutation" and a random hill-climber for efficient numerical tuning, showing that dedicated operators outperform standard GP methods for this task.
- Computational Cost Management: The computational expense of fitness evaluation, especially with FEM, is a major bottleneck. The paper demonstrates the effectiveness of iterative fitness refinement (gradually increasing precision) as a strategy to manage this cost and guide evolution.
- Importance of Constraints: For physical models, incorporating domain knowledge as constraints (e.g., ensuring valid strain energy functions) is crucial. The paper shows that a priori discarding invalid models and using a posteriori checks significantly improves the efficiency and success rate of GP.
- GP's Advantage with Derivatives: The ability of GP to directly handle and evolve expressions involving derivatives is highlighted as a key advantage for mechanical modeling problems where derivatives are integral to the underlying physics and numerical solvers.
Key Results
- Rheological Models: GP successfully identified known rheological models for polyethylene, achieving results comparable to or better than blind random search, with a reasonable success rate (20% for the random hill-climber with stubbornness 2). The iterative fitness refinement strategy showed promise in reducing computation time.
- Strain Energy Functions: GP was able to identify the linear form of the Mooney-Rivlin strain energy function, achieving very low errors on the displacement predictions. The paper also showed that GP outperformed a stochastic iterated hill-climber in terms of error on displacements, with the best GP runs finding accurate coefficients for the $W_g$ function.
Limitations and Future Work
- Computational Complexity: The primary limitation remains the high computational cost of fitness evaluation, especially for the 3D hyperelasticity problem due to the FEM solver.
- Iterative Scheme Refinement: While iterative fitness refinement was effective, the paper notes that more sophisticated switching strategies are needed to prevent populations from getting trapped in local optima of intermediate fitnesses.
- Tree Simplification: The complexity of evolved trees can increase computational cost. Future work could involve developing operators for syntactic simplification of trees.
- Multi-Objective Optimization: The paper suggests exploring multi-objective GP approaches, treating different fitness cases as separate objectives, to characterize Pareto optimal solutions.
- Real-World Data: The current work relies on simulated data. Future work would involve applying these methods to actual experimental data.
- Robustness of FEM Solver: The paper acknowledges that the Newton method used in the FEM solver can fail to converge for complex or poorly conditioned strain energy functions, leading to null fitness. Strategies to address this, such as a priori detection of non-convergent situations, are suggested.
Summary updated