Physics-Informed ML for Semiconductor Epitaxy Control
bravo_abad · x · 2026-07-17
What the Research Did
This work proposes a physics-informed machine learning platform for semiconductor epitaxial growth called SemiEpi. The system reads real-time RHEED video during Molecular Beam Epitaxy (MBE) to decide online whether to increase temperature, maintain parameters, or close the growth window, achieving closed-loop control within a single growth run.
Methodology
The authors split the problem into two parts:
- Using physical relationships to map target emission wavelengths to quantum dot sizes, then using CatBoost to learn the complex nonlinear relationship between size and surface density.
- For real-time control, feeding 24 consecutive RHEED images into a network with global attention residual blocks and cross-layer adaptive fusion modules, specifically designed to handle feature drift caused by rotating substrates.
Results
The paper introduces three classifiers responsible for: surface reconstruction state initialization, detecting if the temperature is too high/low, and when to close the shutter. Validation accuracies reached 99.1%, 99.6%, and 99.9% respectively.
Experimental results showed samples approached the target quantum dot density of 5×10^10 cm^-2, photoluminescence intensity increased by 1.6x, and the linewidth was 29.13 meV. The authors emphasized that this method transfers across reactors because it relies on physical transitions rather than hardcoded device parameters.
More from Infra
- China’s AI arms race is increasingly defined by chips, data centers, and open models — BenBajarin · 2026-07-22
- Agent search bottlenecks are now about variance, not raw latency — rohanpaul_ai · 2026-07-22
- Gavin Baker argues Nvidia may be one of open source AI’s biggest supporters — GavinSBaker · 2026-07-22
- AI Power Demand Exposes US Energy Gap, Urging Shift from Scarcity to Abundance — bradneuberg · 2026-07-22
- Gavin Baker says Nvidia’s $630B figure would be system revenue, not all Nvidia’s — GavinSBaker · 2026-07-22
- A Firecracker-based platform says it can host 6,000 AI agents on one 256 GB server — maritime_sh · 2026-07-22