Advanced Manufacturing AI is where computational science becomes part of the operating manufacturing system. We develop digital twins that combine multiscale simulation, experimental data, in-situ sensing, and reduced-order models into a physics-grounded representation of materials, processes, and equipment. On this foundation, Physical AI can estimate hidden states, interpret mechanisms, predict outcomes, recommend decisions, and eventually interact with the physical process. Our work progresses from virtual evaluation and operator support toward closed-loop virtual–physical integration.
Physical AI requires an internal representation of both observable and hidden process states. We combine multiscale simulation, domain knowledge, and experimental data into world models that connect equipment settings, transport and plasma states, surface chemistry, feature evolution, and manufacturing outcomes. AI accelerates and updates these models without discarding their physical structure.
A digital twin is an executable virtual counterpart of a materials–process–equipment system. Surrogate and reduced-order models make high-fidelity simulations fast enough for repeated evaluation, while virtual metrology estimates quantities that cannot be measured directly, such as particle fluxes, energy delivery, surface states, and evolving feature profiles.
Physical measurements keep the virtual model connected to the operating process. OES, QMS, and equipment signals are analyzed with machine-learning models and fused with simulation results to estimate process state, etch rate, endpoint, drift, and anomalies. These observations also provide evidence for updating and validating the digital twin.
The digital twin provides a safe environment for comparing recipes and equipment conditions before physical execution. Surrogate models and optimization algorithms search the operating space, while sensitivity analysis, SHAP attribution, and uncertainty estimates explain why a condition is recommended and how reliable the prediction is. This establishes a path from offline virtual design to predictive process control.
Agentic systems coordinate simulation, databases, sensing, optimization, and human expertise as one workflow. The long-term loop is to see, understand and predict, decide, and act, with every decision grounded in process physics and checked against the physical system. We are developing this progressively, beginning with engineering decision support and moving toward human-supervised autonomous manufacturing.
DFT
MD
MLFF
DSMC
CFD
Surrogate and reduced-order models
Simulation databases
Experimental databases
OES
QMS
Equipment signals
CNN / MLP
XAI
SHAP
Bayesian optimization
Uncertainty quantification
Digital-twin synchronization
Agent orchestration
Human-in-the-loop and control interfaces
Cryogenic and high-aspect-ratio etching
Virtual metrology for plasma processing
Deposition and reactor-system extension