Research map / Probabilistic and scientific modeling / Molecules and proteins

SimFair: Physics-Guided Fairness-Aware Learning with Simulation Models

AAAI 2024 · Probabilistic and scientific modeling › Molecules and proteins · approach: Counterfactual and statistical explanations · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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.

Proposes SimFair, a physics-guided fairness-aware learning framework that uses inverse-modeled physics simulators (via invertible bijector networks), a dual-fairness consistency loss, and physical-rule constraints to improve location-based fairness of temperature prediction in unlabeled test regions.

Problem

Location-based fairness of satellite-based temperature prediction degrades in new regions and time periods where no labels are available, and purely data-driven fairness methods transfer poorly under distribution shift and changing location groups.

Approach

Use physics-based mechanistic simulators (CMEM, MODTRAN) inverted by bijector-based invertible networks to produce preliminary label estimates in the test region, then align the direction of prediction errors with simulation errors (dual-fairness consistency) and add physical-rule losses so fairness gradients better reflect true fairness.

Key results

• On AT1 with FNN, SimFair-P reduces fairness loss on West-East split to 1.89 vs 4.27-5.56 for baselines (e.g., BaseNet 4.49).
• On AT1 temperature-zone Hot-Cold split, SimFair-P obtains fairness 4.43 vs BaseNet 12.36, while RMSE drops from 19.7 to 12.37.
• Inverse approximation via invertible network reduces RMSE versus directly swapping inputs and outputs for CMEM, per the paper's claim.

Measured effects reported in the paper

Surprising findings

• Several baselines (e.g., Self-Reg, RegFair) sometimes worsen fairness relative to BaseNet on the East-West split, while SimFair remains consistent across splits.
• SimFair is reported to improve RMSE as well as fairness, which the authors attribute to regularization from the invertible structure.

Limitations

• Evaluated only on temperature prediction with two physics simulators (CMEM, MODTRAN).
• Preliminary fairness from simulated labels is not guaranteed to reflect true fairness, as the authors acknowledge.
• Results aggregated over only 5 runs; some baseline variances overlap with SimFair results.

Experimental setup

Datasets: USCRN station air temperature (AT1) SURFRAD air temperature (AT2) SURFRAD land surface temperature with Landsat and MODTRAN inputs (LST) ERA5 hourly reanalysis AMSR2 satellite observations NCEP Reanalysis ASTER Global Emissivity

Benchmarks: AT1 (USCRN-CMEM) AT2 (SURFRAD-CMEM) LST (SURFRAD-MODTRAN)

Metrics: RMSE correlation coefficient location-based fairness

Baselines: BaseNet Sim SimPhy RegFair Self-Reg

Builds on: normalizing flows (bijectors) SimPhy RegFair

Closest papers by meaning

• Towards Harmless Rawlsian Fairness Regardless of Demographic Prior (NeurIPS 2024 · similarity 0.85)
• LoGoFair: Post-Processing for Local and Global Fairness in Federated Learning (AAAI 2025 · similarity 0.83)
• A Local Method for Satisfying Interventional Fairness with Partially Known Causal Graphs (NeurIPS 2024 · similarity 0.83)
• Interventional Fairness on Partially Known Causal Graphs: A Constrained Optimization Approach (ICLR 2024 · similarity 0.82)
• Is Your Model Fairly Certain? Uncertainty-Aware Fairness Evaluation for LLMs (ICML 2025 · similarity 0.82)
• FairPFN: A Tabular Foundation Model for Causal Fairness (ICML 2025 · similarity 0.82)