In high-temperature manufacturing, operators must reach a target state \((T^*,P^*)\) under safety, defect-risk, and energy constraints, while accounting for the fact that once active pressurization stops, continued heating induces thermal expansion that passively raises pressure. Thus, the most critical and low-cost decision is when to stop pressurization: ideally, \((T^*,P^*)\) is achieved by heating alone without over- or under-pressurization; however, rule-of-thumb decisions are difficult to audit and often unstable. We propose an interpretable deep learning approach that reduces reliance on operator experience. Using a small, physics-aligned feature set—current temperature/pressure, recent temperature slope \(k_T=\textrm{d}T/\textrm{d}t\) [ \(^{\circ}\text {Cs}^{-1}\) ], and gaps to \((T^*,P^*)\) —we train a lightweight soft sensor to predict the expansion-phase pressure-rise slope \(\hat{k}_P\) [ \(\hbox {kPa}\,\hbox {s}^{-1}\) ] after stopping. Coupled with a three-phase efficiency model (slower just after stopping, faster in the mid phase, slower near the target), we compute an effective time \(\Delta t_{\textrm{eff}}\) to extrapolate pressure at \(T^*\) and trigger stopping via a transparent “tolerance + M-consecutive-frames” rule. We evaluate on industrial data with train/validation/test splits by run ID. In replay analyses, the same-actual-stop evaluation achieves 93.75% coverage (15/16 runs) under a 5% relative tolerance, with a mean absolute pressure error of \(2.44\pm 2.09\) kPa and a mean absolute percentage error of \(2.03\pm 1.49\%\) . At the earliest stable stop, the method achieves a 100% hit rate over 11 eligible runs, with \(|P(T^*)-P^*|=1.08\pm 0.65\) kPa and a median cycle-time gain of 120 s. The framework is lightweight and auditable, supports de-experientialized stop timing, quantifies cycle-time benefits, and provides uncertainty-aware guidance for tolerance setting.
