Research map / Probabilistic and scientific modeling / Molecules and proteins
End-to-End Phase Field Model Discovery Combining Experimentation, Crowdsourcing, Simulation and Learning
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.
Presents PHASE-FIELD-LAB, an integrated platform combining an AI-assisted superpixel annotation tool, a constrained neural PDE learning module (Neuradiff extended with expert range penalties), and simulation/visualization for discovering phase-field models of nano-void evolution from in situ experiment videos.
Problem
Learning phase-field physics models from large volumes of in situ experimental video data is bottlenecked by costly pixelwise annotation, black-box learners that violate known physical parameter constraints, and a lack of tools for experts to qualitatively assess learned models.
Approach
Combines a superpixel-based click-and-erase annotation tool with a Neuradiff-style recognition-plus-neural-PDE network whose loss adds penalties for violating expert-specified parameter ranges, wrapped in a web interface with simulation and visualization for end-to-end discovery.
Key results
- Annotation tool cut annotation time by 49.8% (circular voids, 282s vs 558s per frame) and 74.7% (mixed voids, 411s vs 1637s) versus VIA, with higher IOU (75.2% vs 53.8% circular).
- Learning module matched vanilla Neuradiff in pixelwise annotation accuracy (~96%) and simulation MSE (~3.3e-4 per pixel) while always satisfying expert parameter ranges, which baseline violated on multiple runs.
- Platform supported discovery of nano-void size fluctuation under Kr irradiation of Cu, with 1% of frames of an 8-minute video annotated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PHASE-FIELD-LAB instead of VIA for annotation | nano-void annotation of circular voids · Circular voids | annotation time (s) per video frame: 558 ± 60 -> 282 ± 38 | Helped |
| Use PHASE-FIELD-LAB instead of VIA for annotation | nano-void annotation of circular voids · Circular voids | IOU (%): 53.8±2.86 -> 75.2±2.82 | Helped |
| Use PHASE-FIELD-LAB instead of VIA for annotation | nano-void annotation of elongated and circular voids · Elongated and circular voids | annotation time (s) per video frame: 1637 ± 202 -> 411 ± 47 | Helped |
| Use PHASE-FIELD-LAB instead of VIA for annotation | nano-void annotation of elongated and circular voids · Elongated and circular voids | IOU (%): 39.1±7.39 -> 74.4±3.78 | Helped |
| Add expert-provided parameter-range penalties to the Neuradiff loss | phase-field model learning · synthetic void-evolution video | satisfaction of known parameter-range constraints | Helped |
| Add expert-provided parameter-range penalties to the Neuradiff loss | phase-field model learning · unseen video frames | pixelwise accuracy: ≈96% vs ≈96% | No clear effect |
| Add expert-provided parameter-range penalties to the Neuradiff loss | phase-field model learning · 100 timestep simulation | MSE per pixel: ≈3.3 × 10−4 vs ≈3.3 × 10−4 | No clear effect |
Surprising findings
- Vanilla Neuradiff-trained models violated at least one of 9 known parameter-range constraints on multiple runs despite similar accuracy.
Limitations
- Evaluation of the learning module uses only a single synthetic video with two voids
- User study limited to 10 male graduate students and two frames
- Performance of both annotation tools degraded with noise in video frames
Experimental setup
Datasets: in situ TEM radiation experiment videos of irradiated Cu synthetic 128x128 phase-field video
Benchmarks: user study on two 1024x1024 in situ video frames (10 participants) synthetic phase-field video (128x128)
Metrics: annotation time per frame IOU (Jaccard) pixelwise accuracy MSE after 100 timesteps parameter range constraint satisfaction Cohen's d
Baselines: VGG Image Annotation (VIA) vanilla Neuradiff
Builds on: Neuradiff SLIC superpixels VGG Image Annotation Cahn-Hilliard and Allen-Cahn phase field equations
Closest papers by meaning
- Cascade of phase transitions in the training of energy-based models (NeurIPS 2024 · similarity 0.84)
- Vector Field Oriented Diffusion Model for Crystal Material Generation (AAAI 2024 · similarity 0.84)
- AutoSciLab: A Self-Driving Laboratory for Interpretable Scientific Discovery (AAAI 2025 · similarity 0.84)
- Mimicking the Physicist's Eye: A VLM-centric Approach for Physics Formula Discovery (NeurIPS 2026 · similarity 0.83)
- Scalable Diffusion for Materials Generation (ICLR 2024 · similarity 0.83)
- MADE: Benchmark Environments for Closed-Loop Materials Discovery (ICML 2026 · similarity 0.82)
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.