AMI
Additive Manufacturing Intelligence

From print settings to part performance

Author: Huilong (Max) Fu

AMI is a web app that learns how additive-manufacturing process settings drive the properties of printed parts, and predicts them with their uncertainty: pre-trained on published data, refined with in-house tests, then used to explore and optimize a process before anything is printed.

Gaussian-process regressionUncertainty quantification Sobol sensitivityOptimization12 AM methods
Reachy Mini, the open-source desktop robot, rendered from its CAD files, beside a twin whose printed parts are coloured by an illustrative predicted field with contour lines
Reachy Mini and a twin whose printed parts carry an illustrative predicted field, not an AMI result. Reachy Mini CAD © Pollen Robotics / Hugging Face, licensed CC BY-NC-SA. Rendered in Blender for AMI, the twin recoloured; this image is shared under the same licence.

From a part to its tool paths

Reachy Mini's face plate, sliced by CuraEngine at 0.2 mm layers with two walls, 15 % infill and supports under the dome, then played back as a slicer previews it: the first layer slowly, the rest as a time-lapse.

  • Outer wall
  • Inner wall
  • Top and bottom skin
  • Infill
  • Support and skirt

Reachy Mini CAD © Pollen Robotics / Hugging Face, licensed CC BY-NC-SA. Sliced and rendered in Blender for AMI; this video is shared under the same licence.

Workflow

Five stages, one project

Every project runs the same five stages and can be saved, reopened and reported.

1

Define Workflow

The material, the test coupon, the AM method and its parameters, and the KPIs: tensile strength, mass, cost and eight more planned.

2

Generate Prior Data

Published data, found and extracted by AI agents, and in-house test results, with simulations planned: the priors the models learn from.

3

Train AI Models

Load a dataset, choose a model, pre-train an ensemble of Gaussian-process regressors, fine-tune, and test on held-out rows.

4

Deploy AI Models

Predictions with intervals, the spread against measurements, which settings matter most, and the settings that do best.

5

Write Report

Pick the results to keep and build a PDF report of the project.

Processes

Twelve AM methods, four material classes

Material extrusion

FFF / FDM · DIW (robocasting)
PolymersCompositesCeramics

Vat photopolymerization

SLA · Ceramic DLP
PolymersCeramics

Powder bed fusion

SLS · MJF · L-PBF · EB-PBF
PolymersCompositesMetals & alloys

Binder jetting

Binder jetting, then sintering
Metals & alloysCeramics

Directed energy deposition

DED: laser, beam or arc (WAAM)
Metals & alloys

Material jetting

Material jetting of photopolymers
Polymers

Sheet lamination

Ultrasonic AM of metal foils
Metals & alloysComposites
Version 0.1

Demo3: CF/PEEK by FFF

Tensile strength from nine print settings

  • Nozzle diameter and temperature, printing speed, raster angle, infill, specimen and layer thickness, bed temperature and carbon-fibre fraction.
  • Learned from 232 tensile coupons in 23 published studies of PEEK and CF/PEEK.
  • An ensemble of 100 Gaussian-process regressors, the best picked on validation rows, tested on held-out coupons.
  • Deployed for prediction with intervals, uncertainty, sensitivity and optimization.