Development of interpretable, data-driven plasticity models with symbolic regression
Introduction
Though occasionally being criticized as less axiomatic than balance laws (e.g., mass, momentum), the development of constitutive models has helped enable the use of new materials and alloys. Specifically, elastoplastic constitutive models based on Tresca [1], Mises [2], and Hill [3] yield functions have stood as the bedrock for engineering analyses over the years. These three models, specifically focusing on plastic deformation, have reached this status because they posses two qualities: (1) they are accurate and tunable enough to be predictive, and (2) they are interpretable, allowing for good judgment in application and extension.
In contrast to the above constitutive models, which are rooted in physics and extended to fit observations, there has been a trend toward data-driven constitutive model development. Generally speaking these models embrace the notion that accuracy and predictive capability are more important then having a sound understanding of the physics involved. Neural network based constitutive models [4], [5], [6] are one example of this paradigm, wherein a neural network model is trained to predict a response (e.g., stress) as a function of load (e.g., deformation). The resulting neural net is often referred to as a black-box owing to its lack of physical insightfullness and interpretability. Another extreme of the data-driven paradigm is the recent work in data-driven computational mechanics which removes the idea of constitutive models altogether and use data directly for the constitutive relationship [7], [8], [9]. This method, by definition, is a black-box and offers no more insight into the constitutive relation than the underlying data itself.
There have also been examples of data-driven constitutive models that fall somewhere in the middle of the interpretablity spectrum. For instance, the work of Soare et al. [10], Soare and Barlat [11] that uses relatively opaque, higher-order polynomials as a plasticity model; however, the derivation and fitting of the coefficients preserves a well known aspect of plastic yield functions: convexity. Another example of partially interpretable constitutive models are based on hybrid modeling [e.g. 12], wherein the modeling error of a classical physics-based model is corrected using a black-box model.
Physics-informed machine learning is another paradigm that combines interpretable aspects of physics with black-box modeling [13], [14], [15]. In this variant of machine learning, physics are incorporated into the objective function for training of a black-box model; either in the form of a residual of a partial differential equation [14] or the weak form of the equivalent boundary value problem [15]. Because the approach in the current work also uses physics in training, it can be viewed as a physics-informed machine learning method; however, there are two key differences between the current work and other physics-informed machine learning works. Firstly, in contrast to hidden physics models [13] where coefficients in a fixed equation are found, a model is sought with arbitrary functional form. And secondly, the models that are produced in the current work are interpretable (in the form of relatively simple equations).
In many – if not most – cases the use of black-box models causes no issues, so long as the model is predictive. There are, however, advantages to interpretable plasticity models. Most importantly, interpretable constitutive models (1) can help drive knowledge discovery (2) may be used more confidently outside the range of their training data and (3) are usually easier to implement within current computational mechanics frameworks.
A common tool for the construction of interpretable, data-driven models is symbolic regression (SR). SR is a procedure for rapidly finding equations that fit a given dataset. Owing to recent advances in scalability, robustness, and efficiency [16], [17], [18], [19], [20], SR has reached maturity such that it can be used in many complex, real-wold applications. For example, Schmidt and Lipson [21] developed a method for performing implicit SR to discover conservation laws from experimental data. The focus of the current paper is the application of SR to the development of interpretable plasticity models. The two primary contributions of the current paper are (1) the formulation of the plasticity modeling problem as an implicit symbolic regression problem and (2) a framework for obtaining the relevant training data from micromechanical simulation.
The topic of using SR in the development of constitutive models has been investigated before. One early example is the work of Schoenauer et al. [22], which developed a three-dimensional hyperelastic model in terms of a strain energy function. The model had limited success for two reasons. First, inserting physical constraints upon the possible strain energy function proved difficult. Second, the presence of real-valued constants within the generated strain energy expressions caused difficulty in the SR optimization. The presence of real-valued constants have been an obstacle for the general SR community [23], [24]. Since the time of Schoenauer et al. [22], techniques have been developed to address the issue, such as restriction of the constants to coefficients of linearly combined SR terms (which allows for regression to obtain best-fit constants) [25], [26] as well as embedded nonlinear optimization of constants [27]. The latter is more computationally intensive but does not restrict the functional form obtained with SR.
For the most part, SR has been used in plasticity model development as a tool for modifying sub-components of a traditionally developed model. Examples include parametrically homogenized constitutive models, which use SR as a means of tying tunable parameters in a yield function to microstructural features [28], [29], [30]. Another example is the work of Versino et al. [31] who used SR in the creation of a flow stress model as part of a Mises () plasticity framework. They worked with relatively sparse experimentally-based datasets, because designing and performing experiments that could capture the needed data was costly and difficult. They were able to overcome the difficulty by supplementing their datasets with synthetic data points which represented their expert knowledge. Nonetheless, generation of significant quantities of relevant experimental data can be a challenge. For some applications there are high data-output experimental techniques [32], [33], [34] to feed data-hungry SR; however, in the current work an alternative approach is taken wherein data for SR is derived from micromechanical finite element (FE) simulations. The primary benefit is that simulation data is usually easier to generate than experimental data. Because the data is more plentiful, more complex relationships can be investigated. The underlying assumption of such an approach is that the micromechanical model accurately represents the material of interest.
In the current work, a framework is developed for using SR in the formulation of plastic yield functions based on the homogenized responses of micromechanical FE models. Details involved in the construction of the micromechanical FE models, along with the synthesis of their response data with SR, are covered in Section 2. The approach is then tested in three test cases (Sections 3 Test cases, 4 Test case results) illustrating its ability to generate interpretable and accurate plasticity models. Section 5 then discusses some of the key aspects that affect the accuracy and robustness of the method. Using the insights produced in Section 5, the method is applied to the problem of porous plasticity in Section 6. The major findings of this work are summarized in Section 7.
Section snippets
Approach
The approach taken in the current work is a type of automatic multi-scale computational homogenization. Multi-scale computational homogenization is characterized by the coarse-graining of the response of a representative volume element (RVE) to a larger scale [e.g., [35], [36], [37]]. It requires that the RVE be well-characterized and that the scale of the RVE is significantly separated from the scale where the homogenized constitutive model is used (see the work of Geers et al. [38] for more
Test cases
The above approach is illustrated in three verification test cases. In all test cases, a single-element FE model with an isotropic elastic–plastic material model is used as a material point evaluator (MPE, see Fig. 2). A known plasticity model is prescribed to the element and then the model is subjected to a mechanical load. The effectiveness of the above approach is gauged by its ability to recover the prescribed plasticity model from the load-response data of the MPE.
The elastic portion of
Test case results
All SR test cases were repeated 20 times to ensure robustness and repeatability. In each SR run the solution populations are evolved until convergence. Convergence is defined as coming within 5% of the fitness value of the target (true) equation. In the more general circumstance where the target equation is not known a priori, convergence would instead need to be defined in a relative manner. In such a case, the use of a train/test split of the data could be used to quantify the performance of
Robustness and efficiency of test cases
Application of the proposed approach for plasticity model development on higher fidelity material RVEs and more complex yield functions will involve increased computational costs. For this reason, it is important to investigate how the robustness and efficiency of the SR-based approach are affected by computational resources and expected model complexity. The following sections include studies of such relationships.
Application to porous plasticity
The above approach is applied to the development of a constitutive model for porous plasticity in this section. This application represents a realistic scenario characterized by (1) the generation of response data from an inhomogeneous RVE, (2) a much more complex system, and (3) a yield function without an expected true result. The results of the current method are compared to two well-known porous plasticity models, the Gurson model [53] and the Gurson-Tvergaard-Needleman (GTN) model [54].
Conclusions
A framework has been presented in this work for the use of symbolic regression (SR) in the development of interpretable, data-driven plasticity models. The framework is a type of automatic computational homogenization wherein load-response data from RVE simulations of a material system are used to develop a homogenized model with SR. The framework was successful in three test cases, delivering equations which are readily interpretable in an automatic fashion. The successful test cases
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was conducted with support of NASA’s Transformative Aeronautics Concepts Program as part of the Convergent Aeronautics Solutions Project and the Transformative Tools and Technology Project.
J.M Emery and J.D. Hochhalter (partially) were supported for this work by Sandia National Laboratories. Sandia National Laboratories is a multimission laboratory managed and operated by National Technology & Engineering Solutions of Sandia, L.L.C., a wholly owned subsidiary of Honeywell
References (55)
- et al.
Neural network constitutive model for rate-dependent materials
Comput Struct
(2006) - et al.
Data-driven computational mechanics
Comput Methods Appl Mech Eng
(2016) - et al.
Jose Vicente Aguado, Emmanuelle Abisset-Chavanne, and Elias Cueto. Data-driven computational plasticity
Proc Eng
(2017) - et al.
Model-free data-driven inelasticity
Comput Methods Appl Mech Eng
(2019) - et al.
On the use of homogeneous polynomials to develop anisotropic yield functions with applications to sheet forming
Int J Plast
(2008) - et al.
Convex polynomial yield functions
J Mech Phys Solids
(2010) - et al.
Hidden physics models: Machine learning of nonlinear partial differential equations
J Comput Phys
(2018) - et al.
Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations
J Comput Phys
(2019) - et al.
An energy approach to the solution of partial differential equations in computational mechanics via machine learning: Concepts, implementation and applications
Comput Methods Appl Mech Eng
(2020) - et al.
Parametrically homogenized constitutive models (phcms) from micromechanical crystal plasticity fe simulations, part i: Sensitivity analysis and parameter identification for titanium alloys
Int J Plast
(2019)
Two-way multi-scaling for predicting fatigue crack nucleation in titanium alloys using parametrically homogenized constitutive models
J Mech Phys Solids
Parametrically homogenized constitutive models (phcms) from micromechanical crystal plasticity fe simulations: Part ii: Thermo-elasto-plastic model with experimental validation for titanium alloys
Int J Plast
Data driven modeling of plastic deformation
Comput Methods Appl Mech Eng
Micromechanical investigation of ductile failure in Al 5083–H116 via 3D unit cell modeling
J Mech Phys Solids
Predicting the ductile failure of Al5083-H116 specimens with a mechanistic model and no free fitting parameters
Int J Solids Struct
Multi-scale computational homogenization: Trends and challenges
J Comput Appl Math
Analysis of the cup-cone fracture in a round tensile bar
Acta Metall
A new model of metal plasticity and fracture with pressure and lode dependence
Int J Plasticity
Mechanik der festen körper im plastisch-deformablen zustand. Nachrichten von der Gesellschaft der Wissenschaften zu Göttingen
Mathematisch-Physikalische Klasse
A theory of the yielding and plastic flow of anisotropic metals
Proc R Soc Lond A
Simulated micromechanical models using artificial neural networks
J Eng Mech
Neural network for constitutive modelling in finite element analysis
Comput Assisted Mech Eng Sci
Hybrid constitutive modeling: data-driven learning of corrections to plasticity models
Int J Mater Form
Continuous hierarchical fair competition model for sustainable innovation in genetic programming
An empirical study of multipopulation genetic programming
Genet Program Evolvable Mach
Coevolution of fitness predictors
IEEE Trans Evol Comput
Cited by (40)
Sparsity promoting algorithm for identification of nonlinear dynamic system based on Unscented Kalman Filter using novel selective thresholding and penalty-based model selection
2024, Mechanical Systems and Signal ProcessingDiscovering interpretable elastoplasticity models via the neural polynomial method enabled symbolic regressions
2024, Computer Methods in Applied Mechanics and EngineeringInterpretable machine learning for microstructure-dependent models of fatigue indicator parameters
2024, International Journal of FatigueComplementing a continuum thermodynamic approach to constitutive modeling with symbolic regression
2023, Journal of the Mechanics and Physics of SolidsData-driven modelling of the multiaxial yield behaviour of nanoporous metals
2023, International Journal of Mechanical Sciences