Reproduce “Mechanistic Neural Networks for Scientific Machine Learning” on OpenResearch

About this paper

Adeel Pervez, Francesco Locatello, Efstratios Gavves

Mechanistic Neural Networks (MNNs) learn explicit governing differential equations as symbolic representations for dynamical systems, leveraging a novel, parallel, and differentiable ODE solver called NeuRLP. This approach combines the predictive strength of deep learning with the interpretability of mechanistic models, achieving improved accuracy and efficiency across tasks like equation discovery, PDE solving, and N-body prediction.

Key insights

• Explicitly learning governing differential equations as symbolic representations within a deep learning framework can bridge the gap between predictive power and scientific interpretability.
• Re-framing ODE solving as a differentiable quadratic program allows for massive step and batch parallelism, significantly boosting computational efficiency for scientific machine learning applications.
• Parameterizing ODE coefficients with neural networks enables the discovery of complex, nonlinear, and rational functional forms of physical laws that are beyond the capabilities of traditional sparse regression methods.

Summary by alphaXiv.