The field of scientific modeling and simulation is moving towards a more integrated approach, combining machine learning and traditional mechanistic models to improve interpretability and accuracy. This is evident in the development of new frameworks that unify scientific machine learning and data assimilation, as well as the use of simulations as supervision for neural networks. Additionally, there is a growing interest in learning physical laws directly from data, bypassing the need for explicit mathematical formulations. Noteworthy papers in this area include: Simulation-Grounded Neural Networks, which achieved state-of-the-art results in prediction and inference tasks across various scientific disciplines. Universal Physics Simulation, which presented a foundational AI model for universal physics simulation that learns physical laws directly from boundary-condition data. Other notable developments include the use of physics-based Application-Specific Integrated Circuits (ASICs) to harness intrinsic physical dynamics for computation, and the application of simulation-based generative models for robust inference of dynamic systems.
Advances in Scientific Modeling and Simulation
Sources
Modeling Partially Observed Nonlinear Dynamical Systems and Efficient Data Assimilation via Discrete-Time Conditional Gaussian Koopman Network
Simulation as Supervision: Mechanistic Pretraining for Scientific Discovery
Learning Koopman Models From Data Under General Noise Conditions
Bayesian dictionary learning estimation of cell membrane permeability from surface pH data
Universal Physics Simulation: A Foundational Diffusion Approach
Solving the compute crisis with physics-based ASICs
A Simple Approximate Bayesian Inference Neural Surrogate for Stochastic Petri Net Models
Learning from Imperfect Data: Robust Inference of Dynamic Systems using Simulation-based Generative Model