Special issue of computational mechanics on machine learning theories, modeling, and applications to computational materials science, additive manufacturing, mechanics of materials, design and optimization

Wing Kam Liu, Miguel A. Bessa, Francisco Chinesta, Shaofan Li, Nathaniel TraskComputational 2023 · DOI · OpenAlex

This editorial, penned by leading experts in the field, introduces a landmark three-volume special issue of Computational Mechanics. It outlines the strategic integration of Machine Learning (ML) with classical mechanics to create "Scientific Machine Learning" (SciML)—a framework that moves beyond black-box models to leverage centuries of physical knowledge for engineering breakthroughs.

TL;DR

This paper introduces a comprehensive framework for "Scientific Machine Learning" (SciML), which embeds physical laws into data-driven models to solve complex engineering problems. By bridging the gap between "small data" in labs and the massive computational needs of multiscale modeling, these approaches enable real-time digital twins and high-fidelity simulations for materials science and manufacturing.

What Problem Does This Paper Address?

Traditional Machine Learning (ML) thrives on "big data" (like millions of internet images), but scientific engineering faces a "small data" problem. Experiments are expensive, and high-fidelity simulations are computationally taxing. Furthermore, standard ML models are "black boxes" that often ignore fundamental physical laws, such as the conservation of mass or energy.

The authors address the challenge of how to efficiently navigate heterogeneous, multimodal data while leveraging existing mathematical and physical knowledge. The goal is to move away from using data to "relearn" known laws and instead use it to fill gaps in our understanding, create faster surrogate models, and solve "unsolvable" ill-posed problems where information is incomplete.

Methods

As an editorial for a special issue, the paper categorizes a wide array of novel methodologies across three volumes:

  • Scientific Machine Learning (SciML): Techniques that augment traditional data-driven models with physical and mathematical domain expertise.
  • Reduced-Order Modeling (ROM) & Surrogates: Constructing efficient models that can mimic expensive simulations at a fraction of the computational cost (often several orders of magnitude faster).
  • Physics-Constrained Learning: Using neural differential operators and physics-informed neural networks (PINNs) to ensure ML outputs remain physically consistent.
  • Multiscale Integration: Methods for "homogenization," which allow researchers to predict macro-scale material properties based on micro-scale or even nano-scale data.
  • Digital Twins: Hybrid frameworks that combine physics-based models with real-time data streams for diagnostics, control, and decision-making in manufacturing.

Datasets

The editorial does not list specific shared datasets, as it summarizes a collection of diverse research papers. However, it highlights the types of data involved:

  • Simulation Data: Direct numerical simulations used to train surrogate models.
  • Experimental Data: "Small data" footprints from large-scale user facilities.
  • Multimodal Data: Combining text, images (e.g., from 3D printing processes), and sensor time-series.
  • Specific Applications: Data from fuel cell modeling, 3D printing thermal compensation, and lattice discrete particle models.

Key Insights and Ideas

  • Don't Relearn the Wheel: The authors argue it is wasteful to use data to relearn fundamental laws (like Newton’s laws). Instead, hybrid frameworks should leverage existing knowledge to allow for better interpolation between limited experimental points.
  • The "Process-Structure-Property" Bridge: AI/ML-enabled digital twins are identified as the critical bridge for manufacturing, allowing researchers to link how a part is made (process) to its internal makeup (structure) and its eventual performance (property).
  • Mathematical Synergy: There is a deep, exploitable mathematical symmetry between topological/geometric machine learning and the theories underpinning physical simulations.
  • Solving the "Unsolvable": ML is uniquely suited for "ill-posed" problems—scenarios where traditional models fail because the information is incomplete or the exact physical form is unknown.

Key Results

While the editorial summarizes many papers, it highlights several high-level quantitative and qualitative impacts:

  • Speed: ML-enabled predictive surrogates are noted to be several orders of magnitude faster than traditional forward simulations, enabling real-time characterization.
  • Efficiency: SciML tools significantly improve data efficiency, requiring fewer experimental data points to achieve high accuracy.
  • Accuracy: By assimilating additional data into traditional models, researchers have achieved higher accuracy in regimes where traditional methods perform poorly.
  • Control: The integration of reinforcement and active learning has provided new tools for the optimal control of mechanical systems and the design of experiments.

Limitations and Future Work

The authors identify several frontiers that still need to be addressed:

  • Interpretability: There is an ongoing need to better understand and interpret why an ML model makes a specific prediction, primarily through physics-informed strategies.
  • System-Scale Integration: While surrogate models for individual components are maturing, integrating these into holistic, system-scale models for complex engineering systems remains a challenge.
  • Uncertainty Quantification: Improving how models handle geometric and probabilistic uncertainty is a key focus for future digital twin development.
  • Generalization: Ensuring that a model trained on one set of differential equations can "transfer" its learning to others (geometry transferability).

Summary updated

BibTeX
@inproceedings{liu2023special,
  author = {Wing Kam Liu and Miguel A. Bessa and Francisco Chinesta and Shaofan Li and Nathaniel Trask},
  title = {{Special issue of computational mechanics on machine learning theories, modeling, and applications to computational materials science, additive manufacturing, mechanics of materials, design and optimization}},
  booktitle = {Computational},
  year = {2023},
  doi = {10.1007/s00466-023-02337-4},
}

Loading related content...

Figure