AI Scientist Platform for research papers in AI, physics, and computational imaging

Computational imaging systems routinely fail in practice because the assumed forward model diverges from the true physics, yet no existing framework systematically diagnoses why reconstruction degrades. We introduce Physics World Models (PWM), a universal diagnostic and correction framework grounded in the Triad Law: every imaging failure decomposes into exactly three root causes -- recoverability loss (Gate 1), carrier-noise budget violation (Gate 2), and operator mismatch (Gate 3). PWM compiles 64 modalities spanning five physical carriers (photons, electrons, spins, acoustic waves, and particles) into a unified OperatorGraph intermediate representation comprising 89 validated operator templates. Autonomous, deterministic agents diagnose the dominant failure gate and correct the forward model without retraining any reconstruction algorithm. Across 7 distinct modalities (9 correction configurations), correction yields improvements ranging from +0.54 dB to +48.25 dB. Gate 3 is identified as the dominant bottleneck in every validated modality, demonstrating that a decade of solver-centric progress has overlooked the principal source of imaging failure.

The computational imaging community has built increasingly powerful reconstruction algorithms, yet real-world deployments routinely fail. We show that a 5-parameter sub-pixel operator mismatch -- well within manufacturing tolerances -- degrades the state-of-the-art CASSI transformer (MST-L) by 13.98 dB, erasing years of algorithmic progress. This paper argues that the bottleneck is not the solver but the infrastructure around it: evaluation protocols, physics representations, calibration pipelines, and benchmarks. Drawing on the SolveEverything.org framework, we present the Physics World Model (PWM) as the "rail" for computational imaging -- a standardized evaluation harness comprising: (i) OperatorGraph intermediate representation (IR), a universal DAG representation spanning 64 modalities across 5 physical carriers with 89 validated templates; (ii) a 4-scenario evaluation protocol separating solver quality from operator fidelity; (iii) the Leaderboard for Imaging Physics (LIP-Arena), a prospective Commit-Measure-Score competition eliminating benchmark overfitting; and (iv) a Red Team adversarial verification module. Across a 26-modality benchmark, operator correction improves reconstruction by +0.54 to +48.25 dB across 9 correction configurations spanning 7 distinct modalities.

Coded aperture snapshot spectral imaging (CASSI) captures a 3D hyperspectral cube from a single 2D measurement using a coded mask and spectral dispersion. Deep learning reconstructors such as MST achieve state-of-the-art quality (>34 dB) but assume perfect knowledge of the forward operator. In practice, sub-pixel mask misalignment and dispersion drift are unavoidable, yet even moderate mismatch degrades MST-L reconstruction by over 16 dB. We propose a two-stage differentiable calibration pipeline: (1) a coarse hierarchical grid search scored by GPU-accelerated GAP-TV, followed by (2) joint gradient refinement through an unrolled differentiable forward operator using a Straight-Through Estimator (STE) for integer dispersion offsets, plus a 1D grid search for dispersion slope recovery. The pipeline is self-supervised, requiring only the measurement and nominal mask -- no ground truth scene. We evaluate five reconstruction methods (GAP-TV, MST-S, MST-L, HDNet, PnP-HSICNN) across four scenarios, revealing a mask-sensitivity spectrum: mask-guided transformers suffer catastrophic degradation (>15 dB) but gain most from calibration (~3 dB).

Compressive imaging faces a critical sim-to-real crisis: models trained on idealized forward operators fail catastrophically when deployed on real hardware. Operator mismatch -- the gap between assumed and true forward operators -- degrades deep learning reconstruction by 10-21 dB, yet no existing benchmark measures this effect. We introduce InverseNet, the first cross-modality benchmark for operator mismatch in compressive imaging, spanning coded aperture snapshot spectral imaging (CASSI), coded aperture compressive temporal imaging (CACTI), and single-pixel camera (SPC). InverseNet evaluates 11 reconstruction methods under a standardized three-scenario protocol -- ideal (I), mismatched (II), and oracle-corrected (III) -- across 27 test scenes and over 240 experiments. We discover an inverse performance-robustness relationship: methods achieving the highest ideal PSNR suffer the largest mismatch degradation -- confirming that a mediocre algorithm with a correct forward model outperforms a state-of-the-art network with a wrong one. All reconstruction arrays, per-scene metrics, and analysis code are publicly released.

Coded aperture snapshot spectral imaging (CASSI) acquires hyperspectral data cubes in a single shot but requires accurate knowledge of the forward measurement operator -- the coded aperture mask position, orientation, and dispersive element parameters -- for high-quality reconstruction. In practice, manufacturing tolerances and assembly drift introduce operator mismatch that degrades reconstruction by 10-17 dB. We present a differentiable calibration framework that models CASSI mismatch as a 6-parameter perturbation (spatial shift, rotation, dispersion slope and axis angle) and recovers these parameters through a two-stage pipeline: (1) a hierarchical beam search over a coarse parameter grid (~38 s/scene), followed by (2) a joint gradient refinement using differentiable PyTorch modules -- including a straight-through estimator for integer dispersion offsets and an unrolled GAP-TV solver with gradient checkpointing (~366 s/scene). Validated on 10 KAIST hyperspectral scenes, our method achieves a calibration gain of +5.06 dB, recovering 30% of the 16.60 dB mismatch loss.