Ace-Mini-1 · Artificial Casting Engineer

A compact, geometry-conditioned research surrogate for 17-4PH gravity investment casting. Two networks contain 197,364 parameters: 166,545 for casting fields and 30,819 for deformation progress.

Try the prediction lab · Interactive solver comparisons

Evaluation and research only. Commercial and production use require separate permission. See LICENSE. This is a restricted research release, not an unrestricted open-source license.

What is included

• model.safetensors: exact V003 field network weights.
• deformation.safetensors: exact V003 deformation network weights.
• config.json: normalization statistics, task/channel order and units.
• Architecture, feature preprocessing, a small inference wrapper and a synthetic example.
• SHA256SUMS: integrity checksums for the release files.

This release does not include training meshes, STEP parts, solver exports, the CAD mesher, the browser app or the private development repository. The released tensors match the deployed V003 checkpoints; removing a geometry from the public examples did not retrain the model.

Quick start

Download this model repository with hf download DigitalMetal/ace-mini-1 --local-dir ace-mini-1, then:

cd ace-mini-1
python -m pip install -r requirements.txt
python example.py

To use CUDA, install a CUDA-compatible PyTorch build and construct AceMini(device="cuda"). The included example only checks numerical execution; it is not a castability validation.

Inputs and coordinate convention

The loader accepts prepared geometry, not a raw STEP file. Use the hosted app for the complete STEP workflow. Callers must rotate geometry, normals and inlet vectors into a right-handed casting frame with gravity [0, 0, -1], using the same translation for surface, query points and gate/inlet locations. Rotate vector outputs back if required.

data: encoder_surface_mm (M×3 surface samples, millimetres), encoder_normals (M×3 outward unit normals), points_m (N×3 query coordinates, metres). Features use the bounding box of the encoder samples and nearest-sample distance as a geometric descriptor; this is not exact wall thickness. The network concatenates max- and mean-pooled point features with spatial Fourier features, gate proximity and inlet direction features.

spec.geometry: attachment_points_mm, inlet_centers_mm, inlet_directions (each G×3, directions pointing with incoming flow), positive gate_radius_mm, and gate_count matching G. spec.process.inlet_velocity_m_s must be 0.5 for the supported process. See example.py for the schema. Surface sampling, CAD repair and meshing affect predictions; the minimal example does not reproduce the hosted app's full CAD preprocessing.

Fixed process: 1,600°C metal, 800°C shell, 6 mm shell; 17-4PH gravity investment casting. These process settings are not freely conditioned model inputs.

Outputs

predict(spec, data, time_s, deformation_phase) returns arrays with channel order specified in config.json:

• maps: filling arrival time and solidification time (seconds), microporosity and total shrinkage (percentage points).
• filling: temperature (Kelvin), solid fraction (0–1), XYZ velocity (m/s), pressure (native solver units; pressure units have not been independently audited), signed liquid-interface distance (metres; negative liquid, positive air).
• cooling: temperature (Kelvin) and solid fraction (0–1).
• mechanical: XYZ displacement and displacement magnitude (metres).
• deformation_m: XYZ displacement (metres) at normalized phase 0–1, with fixed zero/start and final displacement constraints. Phase is not a prediction of physical duration.

Maps and final mechanical fields use time zero as during deployment. Dynamic fields use time in seconds. Query different times for an animation; display interpolation does not create new physical simulation evidence.

Evidence and limitations

Training uses numerical casting-solver results across seven source geometries, with a held-out geometry used in development evaluation. Some showcased comparisons include fitted examples. This is limited solver agreement evidence, not independent physical validation. See the linked comparisons for measured metrics and their definitions; no general accuracy or speed guarantee is made for arbitrary new parts.

Gate counts above two and oblique inlets have limited or unvalidated support. Predictions may violate conservation, produce nonphysical defect values or misrepresent filling/solidification. Stress tensors, room-temperature warp and a physical deformation timeline are not supported. Validate new geometries with an independent solver and appropriate physical testing. Do not use this research model for production manufacturing decisions.