Beyond empirical models: Discovering new constitutive laws in solids with graph-based equation discovery

summary

The gist

Constitutive models are fundamental to solid mechanics and materials science, underpinning the quantitative description and prediction of material responses under diverse loading conditions This work…

In short

The framework uses a graph-based approach to automatically discover constitutive laws from experimental data. It represents equations as directed graphs where nodes are variables and edges encode dependencies, allowing for the discovery of compact, accurate models with material-specific parameters. This method successfully found new strain-rate effects models for steel and lithium metal that outperform traditional empirical methods.

Key concepts

Graph-based Equation Discovery
This technique represents mathematical equations as directed graphs. Nodes represent variables or operators (like 'add' or 'exp'), and edges show the computational relationships between them. This structure allows the system to search for complex, free-form symbolic expressions that describe material behavior, rather than relying on pre-defined equation structures.
Edge Features
In this framework, edge features are used to store information about the parameters in an equation. Instead of having separate nodes for every parameter, these features directly encode the material-specific coefficients that need to be optimized. This makes the graph representation compact and flexible for discovering laws with undetermined constants.
Genetic Programming
Genetic programming is an optimization strategy used to search through possible equation structures. It works by generating candidate graphs based on predefined templates, evaluating them against experimental data, and then evolving the best candidates through mutation and crossover. This process guides the system toward discovering equations that accurately fit the observed physical responses.
Graph Cross-over
Graph crossover is a mechanism used during evolution to recombine successful parts of two parent graphs. It involves randomly selecting nodes in each graph and swapping their corresponding subgraphs. This allows the system to combine functional motifs from different successful models, helping to create more complex and accurate constitutive laws.

Terminology used across episodes

This episode discusses

The paper

Beyond empirical models: Discovering new constitutive laws in solids with graph-based equation discovery · Read on arXiv

Hao Xu, Yuntian Chen, *Dongxiao Zhang

Zhejiang Key Laboratory of Industrial Intelligence and Digital Twin · Department of Electrical Engineering, Tsinghua University · Ningbo Institute of Digital Twin · Institute for Advanced Study, Lingnan University

Constitutive models are fundamental to solid mechanics and materials science, underpinning the quantitative description and prediction of material responses under diverse loading conditions. Traditional phenomenological models, which are derived through empirical fitting, often lack generalizability and rely heavily on expert intuition and predefined functional forms. In this work, we propose a graph-based equation discovery framework for the automated discovery of constitutive laws directly from multisource experimental data. This framework expresses equations as directed graphs, where nodes represent operators and variables, edges denote computational relations, and edge features encode parametric dependencies. This enables the generation and optimization of free-form symbolic expressions with undetermined material-specific parameters. Through the proposed framework, we have discovered new constitutive models for strain-rate effects in alloy steel materials and the deformation behavior of lithium metal. Compared with conventional empirical models, these new models exhibit compact analytical structures and achieve higher accuracy. The proposed graph-based equation discovery framework provides a generalizable and interpretable approach for data-driven scientific modelling, particularly in contexts where traditional empirical formulations are inadequate for representing complex physical phenomena.

DOI: 10.1126/sciadv.aec0989

Transcript

Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.

Kai: We're looking at the paper: "Beyond empirical models".

Mira: Here's the gist of the paper. Constitutive models are fundamental to solid mechanics and materials science, underpinning the quantitative description and prediction of material responses under diverse loading conditions This work proposes a graph-based equation discovery framework for the automated discovery of constitutive laws directly from multisource experimental data Framework Concept The proposed framework expresses equations as directed graphs, where nodes represent operators and variables, edges denote computational relations, and edge features encode parametric dependencies This enables the generation and optimization of free-form symbolic expressions with undetermined material-specific parameters The graph-based representation offers several advantages over conventional tree-based representations, such as allowing node sharing and flexible edge connectivity for more compact encoding Furthermore, the framework embeds material-specific parameters directly as edge features, avoiding the need for additional parameter nodes Optimization Strategy Genetic programming is utilized for optimization through a loop of “generation-evaluation-evolution” The generation process involves introducing graph templates to define structural rules and configurable subgraph patterns Candidate graphs are assembled from randomly generated subgraphs that adhere to these predefined templates, mitigating the risk of generating overly complex structures The evaluation metric for each graph is the mean error computed across all datasets, promoting the discovery of equations with generalizability The L-BFGS optimization algorithm is employed to fit material-specific parameters individually for each material dataset A multi-start optimization strategy is adopted to mitigate sensitivity and retain the solution corresponding to the lowest postoptimization error Key Discoveries The framework has discovered new constitutive models for strain-rate effects in alloy steel materials that exhibit compact analytical structures and higher accuracy compared to conventional empirical models For the deformation of lithium metal, the method identified new constitutive laws directly from experimental data with high accuracy Specifically, a generalized power-exponential form was discovered for the dynamic increase factor (DIF) model This discovered model achieved the lowest mean squared error (MSE) of 6.9×10-4 and the highest coefficient of determination (R2) of 0.980 when compared against classical DIF constitutive models Similarly, for strain hardening, an exponential-form model was selected as the final equation because it provides superior mathematical smoothness and avoids singularities within the plastic strain domain Model Integration and Validation The authors constructed an integrated rate-dependent constitutive relation by combining the independently discovered constitutive models for dynamic increase factor and strain hardening This integrated model achieved a higher accuracy, with an MSE of 646.4 and an R2 of 0.974, reducing the prediction error to nearly half that of the Johnson–Cook (J-C) model The framework demonstrated superior adaptability across different strain-rate regimes for high-strength steel Q460JSC compared to conventional models For lithium metal, the discovered constitutive law showed excellent agreement with experimental data across all temperature regimes, achieving a relative error of only 1.86% The final discovered equation for lithium metal combines a direct linear dependence on stress with a nonstandard evolution term involving an exponential function of temperature T Limitations and Future Directions Several limitations remain, including the specification of graph templates which introduces a form of prior structural bias The current implementation is limited to scalar constitutive responses, and extending the framework to accommodate tensorial quantities and anisotropic material behavior remains an open direction for future investigations Further research is needed to balance representational expressiveness with computational efficiency when dealing with complex problems involving multiple variables or coupled physical fields The work contributes to a new paradigm of scientifically grounded machine learning for physical law discovery Materials and Methods Each candidate equation is encoded as a directed acyclic graph (DAG), denoted by G = (V, E), where V is the set of nodes representing operators or variables, and E is the set of directed edges representing the computational dependencies between them Node types include Variable nodes for input variables like strain rate epsiloṅ or temperature T, and operator nodes drawn from a predefined set such as add, mul, exp, log, and pow Each edge is associated with an edge feature eij ∈ R that represents either a fixed scalar constant or a learnable undetermined coefficient Ci to be optimized during model evaluation Graph templates are introduced to constrain the symbolic search space by defining syntactic and topological constraints on how equations can be assembled Graph crossover is implemented as a subgraph exchange mechanism between two parent graphs to facilitate recombination of functional motifs Graph Evaluation To quantitatively assess the quality of a candidate equation, parameter identification is followed by graph evaluation, which reflects the model’s predictive accuracy with respect to observed data The optimization is performed using the L-BFGS algorithm(51), a quasi-Newton method suitable for smooth objective landscapes and small-to-moderate numbers of parameters The loss function is defined as L(theta) = L(theta) = 1/N ∑ (f(xi; θ) − yi) i=1 to N, where θ = C1 C2 Cn denotes the set of undetermined coefficients encoded in the graph’s edge features The loss is computed as the average loss across all M datasets when data from multiple sources are used, ensuring generalizability Graph Cross-over and Mutation Graph crossover is implemented as a subgraph exchange mechanism between two parent graphs where a random node is selected in each parent graph, and the corresponding subgraphs rooted at these nodes are swapped Three distinct mutation strategies are employed: Edge feature perturbation, Subgraph replacement, and Subgraph deletion Edge feature perturbation modifies a randomly selected edge by replacing its associated feature with a new randomly sampled value Subgraph replacement involves selecting a non-initial node and replacing its subgraph with one newly generated from the graph template Subgraph deletion promotes structural simplification by randomly choosing and removing a subgraph, replacing the parent node with a constant term The evolution process incorporates an extinction strategy where all but a small subset of top-performing graphs are eliminated to maintain population diversity References G.

Kai: A.

Mira: Holzapfel, Nonlinear solid mechanics: a continuum approach for engineering science E.

Kai: A.

Title and authors: Kai: So we're looking at this paper now called "Beyond empirical models: Discovering new constitutive laws in solids with graph-based equation discovery". They are Hao Xu, Yuntian Chen, and Dongxiao Zhang.

Mira: The title really tells you they're not just tweaking existing models; they are trying to discover entirely new ones using this graph structure.

Lev: I wonder how they handle the sheer complexity of finding a structure that is both compact and physically meaningful when you start generating everything randomly?

Kai: They use this graph approach where edges encode the parametric dependencies, which means the material specific parameters are baked right into the connections between operators.

Mira: That’s what’s interesting; they avoid having separate nodes just for those material-specific constants, which is a cleaner way to represent things.

Lev: But that means when you evaluate the graph, you have to be careful about how you're identifying those coefficients so they actually correspond to physical quantities.

Kai: They use L-BFGS optimization with a multi start strategy because they found that the objective function landscape can be tricky, and multiple starting points help them avoid getting stuck in a local minimum.

Mira: The results are pretty compelling because they discovered new constitutive models for strain rate effects in alloy steel materials that have compact analytical structures and higher accuracy than what we see in conventional empirical models.

Lev: Higher accuracy is good, but how does that translate to actual performance on real hardware? Does the complexity of the graph structure translate into a slow evaluation time for something like a real-time simulation?

Kai: That’s where it gets interesting; they found that for the deformation of lithium metal, this method identified new laws directly from experimental data with high accuracy.

The paper's summary: Mira: So to summarize, the core idea is expressing equations as directed graphs where nodes are operators and variables and edges show the computational relations between them.

Kai: The goal is to automate the discovery of constitutive laws by letting a graph generation process create free-form symbolic expressions that have undetermined material specific parameters encoded in the edge features.

Lev: So, instead of starting with an equation structure, you start with templates and randomly assemble subgraphs to form candidate graphs. That sounds like a very smart way to prune the search space initially.

Mira: Right, and then you evaluate each candidate graph by computing the mean error across all datasets from multiple sources to make sure the resulting model is generalizable rather than just overfitting one material.

Kai: They found new constitutive laws for strain rate effects in alloy steel that have compact analytical structures and higher accuracy compared to conventional empirical models.

Lev: That’s a key finding; having a compact structure means it should be easier to implement in simulations, and the accuracy gain is what matters most for us.

Mira: And for lithium metal deformation, they discovered a generalized power exponential form for the dynamic increase factor model which achieved a mean squared error of six point nine times ten to the negative four when compared against classical models <ref:2511.09906#pg3>.

Kai: That MSE number sounds very solid; it’s low enough that it suggests the functional form they found is genuinely capturing some underlying physics in that material system.

Lev: But what about strain hardening, which is often more complex than just a dynamic increase factor? They didn't mention if they found similar high-accuracy laws for plasticity.

The paper's improvements: Mira: One major improvement they highlight is that when you combine the independently discovered constitutive models—for instance, combining their strain rate effect model and their strain hardening model—the integrated relation achieves a higher accuracy with an MSE of six hundred forty-six point four and an R squared of zero point nine seven four.

Kai: That’s significant because they reduced the prediction error to nearly half that of the Johnson Cook model for high-strength steels, which is a benchmark we often compare against.

Lev: Reducing the error by half is substantial; that means their integrated approach provides a much more reliable description of true plastic stress as it depends on both strain and strain rate simultaneously.

Mira: They also showed superior adaptability across different strain rate regimes for the high strength steel Q460JSC compared to conventional models.

Lev: Adaptability is crucial because we're testing this on materials that are undergoing extreme conditions, and if a model works well across different rates, it suggests the underlying physics is being captured more robustly.

Kai: For lithium metal specifically, the final discovered equation combines a direct linear dependence on stress with a nonstandard evolution term involving an exponential function of temperature T.

Mira: That specific combination for lithium metal is interesting because it shows how they handled the coupling between strain rate, temperature, and stress in that system.

Lev: So they managed to find a law that works well across all temperature regimes for lithium metal based on those experimental data points. That's hard to get without significant assumptions about the physics being correct everywhere.

Conclusion: Kai: So, to wrap up this paper, the main thing is that they have successfully used a graph-based equation discovery framework to find new constitutive laws for solids with higher accuracy and better structure than traditional empirical models.

Mira: They achieved this by letting the AI explore free-form symbolic expressions constrained by templates and optimizing those expressions against multiple experimental datasets using genetic programming.

Lev: The implication for us is that if we can build a framework like this, we could automate the discovery of material laws for complex problems without needing to start from scratch every time.

Kai: And they showed that integrating different discovered models—like combining strain rate and strain hardening laws—can lead to a more accurate integrated relation, reducing error compared to existing benchmarks like Johnson Cook.

Mira: For lithium metal, they found a specific law involving stress dependence and an exponential temperature term that fits the experimental data with a relative error of only one point eight six percent across all temperatures <ref:2511.09906#pg1>.

Lev: It’s neat that the framework can handle these complex, coupled responses for materials like lithium metal and high strength steels, even if we still have to worry about how fast the graph evaluation runs on real hardware.

Kai: The paper "Beyond empirical models: Discovering new constitutive laws in solids with graph-based equation discovery" shows a path toward scientific machine learning for finding material science laws.

Mira: It’s a way to move past just fitting curves and start building physically informed, generalizable equations directly from data.

Lev: We'll be watching how this graph structure idea evolves, because if we can constrain the graph search better, it could handle tensorial quantities next, which would be a big step for modeling anisotropic behavior.

More episodes

← Home