Abraham (Chengshuai) Yang

PWM -- Physics World Model

PWM – Physics World Model for Imaging System Autonomy

GitHub Repo

PWM is an open, reproducible evaluation harness and toolkit for computational imaging that aims to make any imaging system self-specifying, self-diagnosing, and self-correcting.

What PWM Does

Given either:

  • A natural-language prompt (e.g., “SIM live-cell, low dose, 9 frames…”)
  • A structured ExperimentSpec (JSON/YAML)
  • Measured data y + an imperfect operator A (operator-correction mode)

PWM produces a fully reproducible run:

Prompt/Spec → OperatorGraph → (Sim or Load) y → (Fit operator θ) → Reconstruct x̂ → TriadReport → RunBundle

Key Features

SolveEverything 10-Gear Framework

PWM implements all 10 gears of the SolveEverything.org abundance engine for computational imaging:

  • 64 imaging modalities covered
  • 89 graph templates for physics operator configurations
  • 43+ reconstruction solvers

64-Modality Coverage

From microscopy to medical imaging to neural rendering:

Domain Modalities
Microscopy Widefield, SIM, Confocal, Lightsheet, STED, PALM/STORM
Medical Imaging CT, MRI, PET, SPECT, Ultrasound, DEXA, Angiography
Spectral Imaging CASSI, CACTI, Hyperspectral, Multispectral
Ultrafast Imaging Compressed Ultrafast Photography (CUP), SCI Video
Electron Microscopy TEM, SEM, EELS
Neural Rendering NeRF, 3D Gaussian Splatting

Physics Fidelity Ladder

4-tier operator hierarchy:

  • Tier 0: Geometry (ray tracing)
  • Tier 1: Wave approximations
  • Tier 2: Full transport/scattering
  • Tier 3: Learned surrogates

4-Scenario Evaluation Protocol

Tests under ideal, assumed, corrected, and oracle mask scenarios with recovery ratio metric.

CT QC Copilot

Clinical medical physics module for diagnostic CT quality assurance.

LIP-Arena

Live Imaging Physics Arena – a built-in adversarial evaluation harness for benchmarking reconstruction methods.

Design Principles

  • Public and extensible – Plugins, CasePacks, dataset adapters
  • Deterministic by default – Bounded search, reproducible seeds
  • Audit-grade – Every run produces a RunBundle with TriadReport, decision records, uncertainty estimates
  • Embeddable into agent systems – Works with AI_Scientist and other agent frameworks

Architecture

PWM is built around the OperatorGraph – a directed acyclic graph of physics primitives that models the entire imaging pipeline from source to sensor. Each node can operate at different physics fidelity tiers.

Python Packages

  • pwm_core – Main library with physics operators, graph IR, reconstruction solvers, calibration, clinical CT QC, viewer
  • pwm_AI_Scientist – Agent bridge for embedding PWM into AI_Scientist

Installation

1
2
pip install -e packages/pwm_core
pip install -e packages/pwm_AI_Scientist

Quick Start

1
2
3
4
pwm run --prompt "widefield deconvolution, low dose, PSF mismatch"
pwm run --prompt "CASSI spectral imaging, 28 bands, coded aperture"
pwm run --prompt "CT sparse view, 90 angles, low dose"
pwm view runs/latest

Academic Papers

PWM includes 5 academic papers with full LaTeX sources:

  1. PWM Flagship Paper (Nature-targeted) – A Universal Physics World Model for Autonomous Diagnosis and Correction of Computational Imaging
  2. InverseNet – Benchmarking inverse problems in computational imaging
  3. PWMI-CASSI – Physics World Model for coded aperture spectral imaging
  4. RAIL – Reconstruction as inference learning
  5. CT QC Copilot / CT QC Platform – Clinical CT quality control

Community

  • 4-level contribution system (solver, calibrator, modality, primitive)
  • Weekly challenges and calibration sprints
  • CISP (Computational Imaging Solver Protocol) challenge for 2026
  • Reputation system and governance framework

Repository Structure

Directory Description
packages/ pwm_core and pwm_AI_Scientist Python packages
papers/ 5 academic papers with LaTeX sources and PDFs
datasets/ InverseNet datasets for CACTI, CASSI, SPC
experiments/ InverseNet, PWM flagship, PWMI CASSI experiments
runs/ RunBundle artifacts across 60+ modalities
tests/ 40+ test files covering clinical, graph, calibration, solver
docs/ Extensive documentation, governance, plans
community/ Contribution guidelines, challenges, leaderboard
rails/ SolveEverything.org 10-gear framework implementation
examples/ 4-level contribution examples, spec files for 16 modalities
results/ Flagship ablation results, statistical tests

Download CVs