The Part Warps When You Fire It

Metal binder jetting — a 3D printing process that jet-prints liquid binder onto metal powder layers, producing "green parts" that must be sintered (heated to near-melting temperature) to fuse into solid metal — promises to be the first additive manufacturing technology economically competitive with casting and machining for medium-volume production runs (hundreds to thousands of parts). Desktop Metal ($1B+ raised, valued at $6.5B post-SPAC) bet on this promise but shut down in 2025 because the sintering step introduces anisotropic, geometry-dependent, and stochastic dimensional distortion that prevents the dimensional precision required for production parts. Parts shrink 15–20% during sintering, but the shrinkage varies by direction, wall thickness, local powder density, and proximity to support structures — making final dimensions unpredictable at the tolerances (±0.1–0.3%) industrial customers require.

Metal additive manufacturing is a $7B+ market growing ~25% annually, but laser powder bed fusion (LPBF) — the dominant technology — is slow (1–10 cm³/hour) and expensive ($300–600/kg for finished parts), limiting it to aerospace and medical applications where geometry complexity justifies the cost. Binder jetting is 10–100× faster than LPBF and uses commodity metal powders (MIM-grade, ~$10–30/kg vs. $50–200/kg for gas-atomized LPBF powders), making it potentially cost-competitive with casting for production quantities. If the sintering distortion problem were solved, metal additive manufacturing could expand from a $7B niche market into mainstream manufacturing for automotive, consumer electronics, and industrial components.

Desktop Metal and ExOne (acquired by Desktop Metal in 2021) developed proprietary simulation software to predict sintering shrinkage and pre-compensate part geometry. However, the simulations rely on material models calibrated to specific powder lots, and real-world powder characteristics vary batch-to-batch (particle size distribution, morphology, oxide content). The sintering process itself is highly nonlinear: shrinkage begins at grain boundaries, proceeds through bulk diffusion, and is influenced by local packing density, binder burnout residues, and thermal gradients within the furnace. Support structure design affects local constraint and distortion but the interactions are complex and geometry-specific. Post-sintering machining can correct distortion but adds cost and defeats the near-net-shape advantage of additive manufacturing. HP Metal Jet (competing binder jetting technology) faces the same sintering challenges and has repeatedly delayed commercial launch. Markforged (Metal X, bound metal extrusion + sintering) entered bankruptcy in 2025 for related reasons.

Physics-informed machine learning models trained on large datasets of sintered part geometries (scan of green part vs. scan of sintered part) could learn the complex, nonlinear mapping between as-printed geometry and final dimensions, enabling accurate pre-compensation without requiring first-principles sintering simulation. In-situ monitoring during sintering (using dilatometry, optical scanning, or acoustic emission) could enable real-time process adjustment. Novel sintering approaches — pressure-assisted sintering (SPS/FAST), microwave sintering, or two-stage sintering profiles — could reduce the temperature and time at which densification occurs, narrowing the distortion window. Standardized characterization protocols for powder feedstock that predict sintering behavior (beyond current particle size distribution specs) would reduce lot-to-lot variability.

A team could build a dataset of sintered vs. pre-sintered part geometries using a desktop binder jetting system (e.g., ExOne Innovent+, widely available in university labs) and train a geometry-to-distortion prediction model. An alternative approach: design a set of benchmark test geometries with varying wall thicknesses, overhangs, and aspect ratios, sinter them, and characterize the anisotropic shrinkage as a function of geometric features. Relevant disciplines: materials science, mechanical engineering, manufacturing engineering, data science.