Knowing When to Stop Sintering: Streaming Detection of the Densification-to-Coarsening Transition

Spark plasma sintering presses a metal or ceramic powder in a graphite die while driving a large pulsed electric current through it, so the powder heats from the inside and becomes a solid part in minutes rather than hours. The difficulty is knowing when to stop. In pure tantalum, a compact held at 1450 °C and 40 MPa for 10 minutes reaches 98.7% of full density but only 340 MPa bending strength. The same conditions held for 2 minutes give 773.8 MPa (Optimizing SPS-Processed Pure Tantalum). The reason is a transition near 96% relative density. Below it, the applied energy closes pores. Above it, the remaining pores stop pinning grain boundaries, grains coarsen from about 4 µm to about 29 µm, and fracture turns brittle. Today the stopping point is fixed in advance by a recipe, and nobody detects the transition while the run is happening.

This proposal treats that transition as a streaming event-detection problem and borrows the problem formulation from a field that has worked on exactly this structure: detecting speech repairs. In speech, the place where a speaker breaks off and restarts (the "interruption point") is a short span whose identity becomes clear only a little after it happens. Researchers there learned two specific lessons. First, scoring whole spans works much better than labeling each word separately. Second, detectors can be trained with little labeled data if pseudo-labels are filtered by an independent plausibility check. We build a bounded-delay span detector over the machine's own logged signals (punch displacement, voltage, current, temperature) and use it to cut the current once the transition is confirmed. The test is whether this beats both fixed recipes and simple change-point rules on parts made from new powder lots.

Core Claim

The densification-to-coarsening transition leaves a detectable signature in standard sintering-machine logs within a short delay. One part of the signature is mechanical: the punch displacement rate decays. The other is electrical: once pores close, the current no longer has to squeeze through narrow particle contacts, which changes the compact's resistance. A detector that scores candidate transition spans using both signal types, with a fixed lookahead window, should locate the transition more reliably than per-timestep classification or a threshold on displacement alone. The reliability gain should be largest when powder lot, particle size, or die geometry shifts the transition away from where a recipe expects it. If that holds, stopping on the detector should keep density at or above 96% while keeping the fine-grained bending strength that the 2-minute tantalum sample showed.

The new element is not the analogy between the two fields. It is a specific detector design carried over from disfluency research, applied to a process-control question that the sintering literature in this file identifies as open and leaves unaddressed. No paper found on Lacuna connects the two fields. The nearest machine-learning work on sintering, Virtual Foundry Graphnet, predicts the final shrinkage and warpage of furnace-sintered parts offline, from geometry. It does not read live process signals and does not make stopping decisions.

Research Program

The method source is the disfluency literature, and each paper plays a distinct role. Transition-Based Disfluency Detection using LSTMs builds an incremental parser that decides word by word whether to keep or delete, reaching 87.5 F1 on Switchboard. It shows that repairs can be found online with a small lookahead buffer, and we reuse that buffered, causal design. Span Classification with Structured Information shows that scoring whole spans rather than single tokens raises Switchboard F1 from 88.2 to 94.1, with the encoder held fixed. Its authors also note that listing every candidate span is too expensive for streaming. We constrain the search so that only spans ending inside the lookahead window are scored, which keeps the cost bounded. Adaptive Unsupervised Self-training trains a student detector on teacher-labeled data. It weights each pseudo-label by whether deleting the candidate span makes the sentence more grammatical, and that check closes most of the gap to supervised training (90.3 against 90.5 F1). We replace the grammar check with a physics check. A candidate transition is kept as a pseudo-label only if the density estimated from displacement before the span follows a normal densification trajectory and the signal after the span does not. Finally, What the Filler? finds that the errors that matter come from structural breaks rather than from easy, salient markers such as "uh". That finding warns us that the obvious cue, a falling displacement rate, may not be where the information lies.

On the sintering side we build a labeled set from interrupted runs on pure tantalum, covering the same grid as the tantalum study: 1350 to 1450 °C and 20 to 40 MPa, at two heating rates and with two powder lots. Each run is stopped at a different time. We then measure the compact's density by water displacement (the Archimedes method) and its grain size by microscopy. Interpolating across stop times gives the time at which density crossed 96% and the time at which grain growth began. About 40 labeled runs are planned. The lab's archive of unlabeled historical logs supplies the data for self-training. The detector's input is the multichannel log, recorded at the machine's native rate. Its output, at each moment, is a decision to stop or continue, made under a fixed lookahead delay. We will vary that delay (2, 5 and 10 seconds) to measure the trade-off between delay and accuracy.

As a portability check on the modified span search, we also run the bounded-lookahead span detector on Switchboard. This tells us whether restricting the span search costs accuracy relative to the offline 94.2 F1 model. That number is a secondary result for streaming disfluency detection. It is not the main claim.

Baselines

The do-nothing reference is the fixed recipe, applied without feedback. We use two settings, 1450 °C / 40 MPa / 2 minutes and 1450 °C / 40 MPa / 10 minutes, both run on the new powder lots. The tantalum study reports 773.8 MPa bending strength for the 2-minute setting and 98.7% density with 340.2 MPa for the 10-minute setting (Optimizing SPS-Processed Pure Tantalum). It does not report density for the 2-minute sample. This baseline is strong because a good recipe may simply be good enough. It also costs nothing, so the detector must beat it on new powder lots to justify itself.

The strongest simple competitor is a rule on the displacement signal. It estimates density from punch displacement after correcting for thermal expansion, then stops when the estimate reaches 96% or when the displacement rate falls below a tuned threshold, whichever comes first. We tune its thresholds by cross-validation on the same labeled runs that train the detector, with the same number of tuning trials. Lacuna has no reported result for any online stopping detector in spark plasma sintering, so this rule is our own construction, not a published number.

Two ablations isolate what the proposal adds. The first is a per-timestep classifier: the same encoder and the same inputs, but each moment is labeled on its own instead of scoring spans. It mirrors the 88.2 F1 token-tagging baseline in the span-classification paper. The second ablation removes the electrical channels, voltage and current, and keeps the mechanical ones. A further ablation trains on the labeled runs only, without self-training. The detector beats the field if, on held-out powder lots, it locates the transition with lower median timing error than the displacement rule and, in closed-loop runs, yields higher bending strength than the 2-minute recipe at density no lower than 96%. On Switchboard, the published reference is the 94.2 F1 offline BERT-GCN model.

Why not just…?

Why not just use the best fixed recipe? If the transition always falls at the same time for a given temperature and pressure, a recipe already captures it, and this project adds nothing. The tantalum results show the transition depends on temperature and dwell time. Whether it also moves with powder lot and particle size is exactly what our held-out-lot experiment measures. If the transition time varies across lots by less than the detector's error, we will report that the recipe suffices.

Why not just threshold the estimated density from displacement? This is the real threat. Displacement-based density estimation is standard practice, and it may track the 96% crossing well. Two things could make it fall short. Thermal expansion and punch-die compliance contaminate the displacement signal. And the onset of coarsening, as opposed to the density crossing, may show up in resistance rather than in displacement. The ablation without electrical channels settles this directly. If the displacement rule matches the full detector, the disfluency-derived design is unnecessary and we will say so.

Why not just build a physics model, such as a finite-element simulation or a learned graph model like Virtual Foundry Graphnet? These models predict the end state from inputs fixed before the run. They do not observe the specific compact as it sinters, so they cannot correct for an unusual powder lot. We use a physics trajectory only as the plausibility filter in self-training, not as the decision rule, and we compare against forward simulation through the recipe baseline.

Proposed Experiments

The first decisive check is offline and uses leave-one-powder-lot-out evaluation. We measure the median absolute error in detected transition time and the rate of premature stops, defined as stops before measured density reaches 96%. Every method in the Baselines section is evaluated at every lookahead delay. If the span detector does not beat the displacement rule by a margin larger than the spread across random seeds, the project stops there, and the negative result is reported.

The second check is closed loop. On about 15 fresh runs from an unseen powder lot, the detector cuts the current live. We measure the resulting density, grain size, three-point bending strength and fracture mode, and compare them with matched runs on the fixed 2-minute and 10-minute recipes. Success means density of at least 96%, grain size close to the fine-grained regime, and bending strength above the fixed-recipe runs on that same lot.

The third check is a transfer test on Ti-15Nb-25Zr-8Fe, the alloy in Effects of Fe on Ti–15Nb–25Zr–xFe Alloys. This alloy has a different failure mode: niobium does not fully dissolve, and brittle iron-titanium and iron-zirconium compounds form. The question is whether the detector, retrained with a handful of labeled runs, still finds a transition, or whether the signature is specific to tantalum.

The last check is the Switchboard run of the bounded-lookahead span detector against the 94.2 F1 offline model. It shows whether the latency-bounded span search is a generally sound design or only a convenience for sintering.