Portfolio
Manufacturing ML / Physics-Informed Few-Shot LPBF Prediction
Manufacturing ML · Process Engineering

A training-free classifier that qualifies unseen LPBF alloys in seconds.

PIKNN screens laser powder bed fusion process windows in an 8-D physics-engineered space. No learnable parameters, no meta-training — a single handbook lookup per alloy. +18.3 pp over a meta-trained Prototypical Network on held-out Ti6Al4V.

  • Physics-informed ML
  • Few-shot (1–10 shot)
  • k-NN
  • scikit-learn
  • PyTorch (baseline)
Dataset
1,579 builds · 6 alloys
Held-out gain
+18.3 pp
Separability
−46% DBI
Train time
0 s vs 11.9 s

J. Intelligent Manufacturing · IISE 2026 Under Review

Physics-informed feature pipeline: five process parameters combined with three derived physics features (energy density, power density, thermal diffusivity) to classify relative density

Why it matters

Bringing a new LPBF alloy from lab to certified aerospace production takes 5–15 years and tens of millions of dollars. A single complex component can demand $4M+ and two years just to qualify under NIST building-block protocols, and every new feedstock or parameter shift restarts the cycle.

Conventional ML approaches need 103–104 labeled experiments per alloy to generalize — effectively rebuilding the dataset every time. But the thermal physics of LPBF is already known before any build runs. If cross-material screening lived in physics rather than data, qualification cycles would collapse from years to days.

NIST building-block qualification pyramid showing four tiers from bottom to top: Coupons (weeks, dollars in thousands, tens of tests), Elements (months, $10K+, hundreds of tests), Sub-components (~year, $100K+, hundreds to thousands of tests), System (5 to 15 years, millions of dollars); a red callout below states that any change among 100+ key parameters resets the process back to coupons
Qualification ladder Any one of 100+ key parameter changes resets the process back to coupon-level testing — where weeks and dollars compound up through elements, sub-components, and system qualification.

How it works

  • Process knobs (ψ1–ψ5). Laser power, scan speed, hatch spacing, layer thickness, powder size — the five primary LPBF parameters.
  • Derived physics (ψ6–ψ7). Volumetric energy density captures melting completeness; power density separates conduction-mode from keyhole-mode regimes.
  • Cross-material bridge (ψ8). Thermal diffusivity αm from a materials handbook carries ~30.6% of squared distance across alloys — the alloy-physics bridge that ProtoNet learns and PIKNN simply reads.
  • Classifier. 1-NN in standardized 8-D space. Zero learnable parameters, no softmax, no gradient updates — vs. ProtoNet's 13,056 parameters and 10,000 meta-training episodes.
  • Validation. 4-way K-shot (K ∈ {1,3,5,10}), 300 episodes per config, 95% bootstrap CIs. Held-out alloys Ti6Al4V and CuCrZr (37× thermal-diffusivity contrast).
Side-by-side architecture: ProtoNet learns an embedding from support and query sets via backpropagation, while PIKNN replaces the entire embedding network with a physics feature mapping plus standardization, then takes a nearest-neighbor decision
Architecture ProtoNet trains an embedding for 10,000 episodes (top). PIKNN swaps that whole network for a physics feature map plus a 1-NN lookup (bottom) — same support/query interface, zero learned weights.
Energy density · Power density ED = P / (v · h · l)  ·  PD = 4P / (πd²)
Dataset composition: pie chart of alloy distribution across 316L, AlSi10Mg, 18Ni300, IN718, Ti6Al4V, CuCrZr; histograms of the seven process parameters; and relative-density distribution with violin plot showing heavy concentration above 95%
Dataset 1,579 builds across six alloys. The two held-out alloys (Ti6Al4V, CuCrZr) sit at opposite ends of the thermal-diffusivity range — a hard cross-material test.

What we found

58.0% vs 39.7% Ti6Al4V · 10-shot PIKNN vs. trained ProtoNet on a held-out alloy — +18.3 pp.
52.0% vs 46.1% CuCrZr · 1-shot +5.9 pp over ProtoNet; +10 pp over a trained SVM.
−46% DBI Class separability Davies–Bouldin: 11.65 (learned) → 6.27 (physics) on held-out union.
0 s training vs. 11.9 s ProtoNet 200 experimental hours of screening compressed to 1.25 s of compute.
Few-shot learning accuracy as a function of K-shot from 1 to 10 for Ti6Al4V and CuCrZr alloys, comparing PIKNN circles solid line, ProtoNet triangles dashed line, and SVM squares dotted line; PIKNN consistently above the baselines especially at higher K
Few-shot gain PIKNN (circles) tracks above ProtoNet (triangles) and SVM (squares) at every K-shot, widening at 5 and 10 shots where physics features generalize cleanly to the unseen alloy.
t-SNE projection contrasting physics features on the left, where relative-density classes form clean colored regions, with ProtoNet embeddings on the right, where the same class colors are mixed and overlap
t-SNE Physics features preserve relative-density neighborhoods across both alloys (left). The learned ProtoNet embedding mixes the classes for the unseen feedstock (right).
Laser powder bed fusion process: schematic of laser, mirror, powder supply and build plate, and a photograph of the build chamber with active laser tracks emitting sparks across a layer of coupons
Context Each coupon in the dataset is one full LPBF print — minutes of build time, hours of post-density measurement. Cutting screening from 103+ to ~10 coupons per alloy is the lever.

So what. Drop in a new feedstock, look up its thermal diffusivity, run 1–10 calibration coupons — get a process-window screen in seconds. Nothing to retrain when alloys or machines change. Applies anywhere physics is known but data is sparse: aerospace alloys, Cu thermal management, high-entropy alloys.

Paper & resources

IISE 2026 talk Presentation slides
Dataset (Dataverse) 10.7910/DVN/VPBQK8
Journal version Under review · J. Intelligent Mfg.
Author Prashant Dhakal
BibTeX · under review 
@article{dhakal2026piknn,
  title   = {Physics-Informed Training-Free Few-Shot Learning for Cross-Material Relative Density Prediction in Laser Powder Bed Fusion},
  author  = {Dhakal, Prashant and Kim, Jae Gwang and Hou, Aolin and Wu, Xiaofei and Wang, Shiren},
  journal = {Journal of Intelligent Manufacturing},
  year    = {2026}, note = {Under review}
}