A Physics-Informed Surrogate Safety Gate for Agentic AI Control of SMT Reflow Soldering in Unmanned-System Electronics Manufacturing: A Simulation-Based Study

Surface mount technology (SMT) assembly lines running under high-mix low-volume (HMLV) conditions face a recurring parameter-optimization problem, because frequent product changeovers invalidate static control recipes and expose the nonlinear coupling between solder paste printing and reflow soldering. A large language model (LLM) can reason about that coupling. Its outputs are probabilistic and its inference latency is network-dependent, so it cannot be wired directly to process actuators. This paper takes that obstacle as its research question. The proposed multilayer Agentic AI and physics-informed digital twin (Agentic AI-DT) framework, a digital-twin architecture evaluated here as a digital model driven by a simulated closed loop, places a cloud-hosted LLM in the strategy role and an edge-deployed conditional variational autoencoder (CVAE), trained with a thermodynamic partial differential equation residual penalty, in the role of a gate that runs the same deterministic check on every strategy before it reaches the reflow oven controller. The check executes deterministically, but its verdicts inherit the surrogate's prediction error, so the safety it provides is measured here and not assumed. The study measures the gate itself: how often it fires, and how often it errs in each direction. A two-stage pipeline combining Latin Hypercube Sampling with singular value decomposition based reduced-order modelling expands 1,000 calibrated finite element simulations into 50,000 training scenarios, and 200 further independent finite element runs form the held-out test set. On those runs the physics-informed CVAE reaches a mean squared error of 1.93 ℃ 2 , 25.2% below a CVAE trained without the physics term and 14.6% below a standalone physics-informed neural network. Across 347 candidate strategies the gate vetoes 63 (18.2%). Finite element re-simulation shows that 5 of the 284 admitted candidates (1.76%, 95% CI 0.8 to 4.1%) were unsafe and that 14 of the 63 vetoed candidates (22.2%, 95% CI 13.7 to 33.9%) were safe, giving 94.5% agreement with the reference solver. Closed-loop profile tracking error is 2.09 ℃, against 3.18 ℃ for model predictive control and 4.62 ℃ for fixed-gain PID. The diagnostic layer names the injected root cause in 84.0% of 300 scripted fault trials, against 72.7% for a random forest classifier and 61.3% for a rule-based expert system. Median end-to-end latency is 419 ms, 90.7% of 2,000 cycles finish inside the 500 ms design budget, and the remaining cycles commit a re-checked last validated setpoint at the 500 ms boundary, reaching the controller within 539 ms. Every number reported here comes from a synthetic environment calibrated against three published sets of measured reflow profiles. No physical production line was instrumented, so these are simulation-fidelity measurements rather than production performance guarantees.