QUALIFICATION SOVEREIGNTY.

• Collaborative robots load and unload specimens on the test frames and move them between stations

• Barcode/ID scanning at each handoff. Every movement is written to the digital thread

• Scheduled queues run lights-out overnight and on weekends, so frames and prep equipment are used for more hours each day

• Load, position and handling are the same for every specimen, which removes operator-to-operator variability

• Small-format build platform for parameter sweeps, density cubes and witness coupons

• Design-of-experiments planning for laser power, scan speed, hatch spacing, layer thickness and contour strategy

• Process window mapping against density, surface and property responses

• Tensile frames for ambient tensile, yield, elongation, reduction of area and modulus (ASTM E8/E8M, ISO 6892-1, ASTM E111)

• Fatigue frames for load-controlled high-cycle fatigue (ASTM E466) and strain-controlled low-cycle fatigue (ASTM E606/E606M)

• Cobot-tended specimen loading, video extensometer and continuous testing in a queue

• Raw load, displacement, strain and cycle data streamed straight into the evidence record

• Automated sectioning, mounting, grinding and polishing (ASTM E3)

• Programmed prep recipes for each alloy, with consumables and step times logged per specimen

• Automated etching (ASTM E407) for repeatable surface preparation

• Motorized optical microscopy with automated stitching of full cross-sections

• Porosity and defect quantification by image analysis (ASTM E1245, ASTM E2109), with gas pores and lack-of-fusion classified separately

• Grain size (ASTM E112, ASTM E1382) and melt-pool and phase features, measured by ML image models

• Automated microhardness traverses (ASTM E384, ASTM E92)

• Laser and optical 3D dimensional inspection with CAD comparison and GD&T reporting

• Areal and profile surface roughness (ISO 25178, ISO 21920 / ASME B46.1)

• First article inspection as the part-qualification gate

Every object in the lab gets an ID and links to everything before and after it:

• Powder lot, chemistry certificate, PSD and reuse history

• Machine, build file, parameter set, scan strategy, plate layout and in-situ build logs

• Coupon and part location, orientation and post-processing history (stress relief, HIP, heat treatment, machining)

• CT volumes, dimensional scans, surface data, micrographs, raw test curves

Together these build an as-built digital twin of each part and coupon, a queryable record of how it was made and how it measured. Any value on a certificate can be traced back through the twin to the powder lot and build behind it.

The models start from physics and add data-driven learning, so they can make useful predictions from the small data sets typical of qualification:

• Process-property prediction. Physics-informed Gaussian process regression predicts density, tensile and fatigue response from process parameters, energy density, geometry and post-processing, with uncertainty bounds on every prediction.

• Parameter optimization. Bayesian optimization recommends the next set of parameter-sweep experiments. This reduces the number of builds and coupons needed to define a process window.

• Defect detection on CT data. Computer vision models segment and classify volumetric defects (gas porosity, lack of fusion, keyhole pores, inclusions) and measure size, shape and location against acceptance criteria.

• Microstructure analysis. Image models measure porosity fraction, grain size and phase features from micrographs, consistently from one operator and one lab to the next.

• Build anomaly detection. Unsupervised models flag unusual behavior in machine logs and in-situ monitoring data before the parts reach the test lab.

• Pass/fail screening. Gradient-boosted classifiers screen parts and coupons once enough labeled results exist. Every prediction is shown alongside the physical features behind it.

• Uncertainty first. Each model reports how confident it is. Low-confidence results go to physical testing instead of being accepted.

Every Deployable Lab is a node in the ASTRO network.

• Models train locally on each lab’s own data.

• Only model updates (learned weights and statistics) are shared with the network. Each customer’s raw data, part geometry and build files stay in that customer’s lab.

• The network model learns process-property relationships across alloys, machines and sites. Each new lab starts with the knowledge of the whole network instead of starting from zero.

• Comparisons across machines and sites show where a process drifts, so variability is caught early.

A large language model trained on qualification work, connected to the lab’s evidence record and the governing standards:

• Qualification planning. Turns a requirement, such as “qualify Inconel 718 on this machine to NASA-STD-6030,” into a draft test matrix, coupon count, build layout and schedule.

• Spec-aware retrieval. Answers plain-language questions against the lab’s own data and the standards behind it, and cites the record or clause each answer came from.

• Data review. Compares measured results with acceptance criteria and points out deviations, outliers and missing evidence.

• Documentation. Drafts test reports, nonconformance reports and release packages for engineers to review.

• Root-cause support. Links a failed coupon to its CT data, fractography, build log and powder lot, and suggests hypotheses for engineers to test.