Process to microstructure: what a 3D printer's laser actually teaches a neural network

Here is a number that annoyed me for a week: 266 K. That’s the temperature swing our surrogate had to track, frame by frame, as a simulated laser crossed a part in a metal 3D print. Not a static thermal field. A moving one.

Most people talking about AI for additive manufacturing show you a picture of a finished part and a color map of “predicted stress.” I want to talk about something less photogenic and more load-bearing: thermal history.

In laser powder bed fusion (LPBF), the part isn’t built, it’s grown, one thin layer at a time, under a laser that sweeps across the powder bed and melts it locally. What ends up inside the metal — the grain structure, whether it’s fine and isotropic or coarse and columnar, whether the part is tough or brittle in a given direction — is decided almost entirely by two derived quantities: the thermal gradient G and the solidification front velocity V at the moment each bit of metal freezes. G and V come out of the thermal history. Get the thermal history right, cheaply, and you have the raw material for predicting microstructure. Get it wrong, and no amount of post-processing saves you.

That’s the part we’ve built and verified so far.

What’s actually running

Additive manufacturing is not bolted onto our simulation engine (NEMO) as a separate tool. It’s one physics mode inside the same unified model that also handles thermal, structural, and other domains — the “additive_mfg” mode in our temporal solver. We ran real, moving-laser transient thermal solves on MOOSE (our finite-element ground truth), and watched the hot spot travel through the part across 7 frames, with the local temperature swinging by that 266 K figure. That’s a genuinely time-dependent problem, not a single static snapshot dressed up as one.

The part I actually care about, from an engineering-honesty standpoint, is how the process parameters get into the model. Laser power density, scan speed, spot radius, and layer thickness aren’t just fed to the physics solver — they’re also written directly into the same conditioning slots (30 through 32) that the neural network reads as input. That sounds like a small detail. It isn’t. It means the number the model sees during training is the exact same number the physics felt during the solve. No re-derivation, no unit drift, no silent mismatch between “what we told the solver” and “what we told the network.” When we varied the power/speed combination (Q, v), the measured peak temperature response moved from 97.7 K to 618 K — a large, clearly causal swing, which tells us the conditioning channel is doing real work, not sitting there as decoration.

The build itself is simulated layer by layer using the quiet-element method: powder sitting above the current build front is modeled at roughly 2% of solid conductivity, so it behaves thermally like loose powder until the laser reaches it and it “activates.” I want to be direct about the limitation here: this is a macro-layer approximation. Real LPBF layers are 30-50 micrometers thick. Our current mesh resolution doesn’t get down there. That’s a genuine gap between the model and the physical process, and I’d rather say so than let a nice chart imply otherwise.

Where the honesty gets uncomfortable

We are integrating two specialist solvers to go further: MALAMUTE, a MOOSE-based tool from Idaho National Laboratory for melt-pool thermo-fluid behavior and pool stability, and AdditiveFOAM, an OpenFOAM-based tool from Oak Ridge National Laboratory (currently v1.2.0) that captures the messier defect physics — spatter, vapor plume, keyhole formation — and produces solidification data suitable for microstructure prediction through ExaCA coupling.

Here is the sentence that matters most in this whole post: we do not have a trained microstructure head, and we do not have a trained G/V head. The mechanism — laser moves, heat flows, process parameters causally shape the thermal response, and that response is captured by a surrogate instead of a full solve — is proven end to end. The prediction pipeline runs. What it doesn’t yet do is tell you, with a validated number, what your grain structure will look like. Until the model goes through co-training with the other physics and passes a held-out accuracy gate, any thermal or microstructure output from this beachhead is mechanism-proven, not numerically trustworthy. Calling it “done” would be exactly the kind of confident-but-wrong claim I try hardest to avoid.

Same caveat applies to the sibling PCB thermal beachhead: 113-board corpus, the pipeline verified 100/100 on synthetic runs, ingest smoked end-to-end on a real 100x140mm board with 86 components — and still no accuracy gate until co-training runs.

The eventual product target is a printability and design report, an operator parameter sheet, and something we call a Build Recipe: a JSON artifact with its own validation block, so that if the numbers behind it haven’t been checked, the recipe is machine-labeled DRAFT rather than silently trusted. Any zone-modulated scan path we generate gets re-verified through the high-fidelity solver before it’s allowed near real hardware.

I’d rather ship that discipline slowly than ship a microstructure map I can’t stand behind.

If you work in LPBF process development, metallurgy, or additive qualification, I’d like to hear where you think G/V-based surrogates break down in practice.