Research map / Optimization / Optimizers
Macrocanonical Generator Networks: data-efficient neural surrogates for amortized physics simulation
Niall Jeffrey, Benjamin Wandelt
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces macrocanonical generator networks, which learn fast, data-efficient neural surrogates for amortized simulation of stationary physical processes by matching multiscale statistics of the target process.
Problem
Scientific simulations are computationally expensive, which limits their use in simulation-based inference, uncertainty quantification, and decision making.
Approach
Learns neural surrogates for amortized simulation of stationary physical processes by matching multiscale statistics of a target process, inspired by microcanonical maximum entropy (generator-network) ideas.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| pretrain SBI encoder on macrocanonical-generated synthetic data | ResNet encoder + Gaussian neural density estimator · Quijote Omega_DE inference | validation NLL: -0.93 -> -1.69 | Improved |
| train from scratch on real simulations instead of synthetic pretraining | SBI encoder · Quijote Omega_DE inference | validation NLL | Hurt |
| remove the residual shortcut (ablation) | macrocanonical generator · MHD and cosmic-web | sample quality / collapse | Hurt |
| use large learning rate eta=1e-2 instead of small (ablation) | macrocanonical generator · MHD and cosmic-web | sample quality / collapse | Hurt |
| replace circular padding with zero padding (ablation) | macrocanonical generator · MHD and cosmic-web | stationarity | Hurt |
| train on more fields (1, 10, 100) (ablation) | macrocanonical generator · MHD turbulence | statistics vs held-out | Improved |
| use macrocanonical training instead of WGAN-GP, flow matching or DDPM | generator · single training field | memorization | Improved |
| match features in expectation instead of per-sample (microcanonical) | generator · Quijote slices | per-realization fluctuations | Improved |
Experimental setup
Builds on: microcanonical maximum entropy
Closest papers by meaning
- Microcanonical Langevin Ensembles: Advancing the Sampling of Bayesian Neural Networks (ICLR 2025 · similarity 0.86)
- Arbitrarily-Conditioned Multi-Functional Diffusion for Multi-Physics Emulation (ICML 2025 · similarity 0.85)
- Learning Stochastic Multiscale Models (NeurIPS 2025 · similarity 0.85)
- Variational Inference with Coverage Guarantees in Simulation-Based Inference (ICML 2024 · similarity 0.85)
- Learning Macroscopic Dynamics from Partial Microscopic Observations (NeurIPS 2024 · similarity 0.85)
- Compute-optimal data scaling for neural surrogates via multi-fidelity training (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
