Hassan Arbabi,Judith E. Bunder,Giovanni Samaey,Anthony J. Roberts,Ioannis G. Kevrekidis
Abstract:The data-driven discovery of partial differential equations (PDEs) consistent with spatiotemporal data is experiencing a rebirth in machine learning research. Training deep neural networks to learn such data-driven partial differential operators requires extensive spatiotemporal data. For learning coarse-scale PDEs from computational fine-scale simulation data, the training data collection process can be prohibitively expensive. We propose to transformatively facilitate this training data collection process by linking machine learning (here, neural networks) with modern multiscale scientific computation (here, equation-free numerics). These equation-free techniques operate over sparse collections of small, appropriately coupled, space-time subdomains ("patches"), parsimoniously producing the required macro-scale training data. Our illustrative example involves the discovery of effective homogenized equations in one and two dimensions, for problems with fine-scale material property variations. The approach holds promise towards making the discovery of accurate, macro-scale effective materials PDE models possible by efficiently summarizing the physics embodied in "the best" fine-scale simulation models available.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: How to efficiently extract macroscopic effective partial differential equations (PDEs) from fine - scale simulation data to describe the macroscopic behavior of materials with microscopic heterogeneity. Specifically, the authors propose a method that combines machine learning (especially neural networks) with modern multi - scale scientific computing (especially equation - free numerical methods), thereby reducing the amount of computation required to collect effective macroscopic - scale PDEs for training.
### Problem Background
In the fields of materials science and engineering, the material properties in many physical or engineering modeling problems vary at a small scale (i.e., heterogeneity), and we are interested in predicting the response of materials at a larger scale. Traditional homogenization techniques derive effective equations describing the large - scale response of materials through asymptotic analysis methods under certain assumptions. However, the scope of application of this method is limited. Especially in the case of non - periodic heterogeneity, it is not possible to systematically derive an effective diffusivity formula. In addition, the computational resources required for direct numerical simulation of heterogeneous materials increase exponentially as the scale ratio \( \epsilon \) decreases, which makes such problems computationally infeasible.
### Paper Solutions
To solve the above problems, the authors propose the following solutions:
1. **Combining Machine Learning and Multi - scale Computation**:
- Utilize equation - free numerics to conduct local fine - scale simulations through sparse space - time sub - domains (called "patch") to generate data for effective macroscopic - scale PDEs for training.
- These methods assume the existence of an effective macroscopic - scale equation but do not explicitly derive it. Instead, they obtain the required macroscopic information through fine - scale simulations within a short time and small - space domain.
2. **Using Neural Networks to Learn Effective PDEs**:
- Learn effective homogenized PDEs through two different neural network architectures:
- The first architecture attempts to discover the functional form of the PDE, i.e., the relationship between time and space derivatives:
\[
\frac{\partial U}{\partial t}=F\left(U, \frac{\partial U}{\partial x}, \frac{\partial^{2} U}{\partial x^{2}}, \ldots\right)
\]
- The second architecture directly models the law of the PDE in discretized form, i.e., relates the time derivative to the field values of neighboring points:
\[
\frac{d U\left(x_{j}\right)}{d t}=G\left(U\left(x_{j}\right), U\left(x_{j - 1}\right), U\left(x_{j + 1}\right), \ldots\right)
\]
3. **Reducing Computational Complexity**:
- Significantly reduce the amount of computation required to generate training data by conducting fine - scale simulations only in part of the space and time domains. For example, in the case of a one - dimensional diffusion problem, using the patch dynamics method can conduct simulations within less than 8% of the space and 0.1% of the time domain, effectively reducing the space - time grid size by about 10,000 times.
### Application Examples
The paper verifies the proposed method through two specific examples:
- **One - Dimensional and Two - Dimensional Diffusion Problems**: Demonstrates how to learn effective homogenized PDEs from fine - scale simulation data and compares them with known analytical solutions.
- **Other Material Problems**: This method can be generalized to other material problems with microscopic heterogeneity, helping scientists more efficiently understand and predict the macroscopic behavior of materials.
Through these methods, the authors hope to develop a practical tool that enables researchers to accurately discover and describe effective macroscopic models of complex material systems under limited computational resources.