One of the central problems in advanced manufacturing is that the same nominal material can exhibit different properties when produced using different manufacturing processes. Powder-bed fusion, directed-energy deposition, casting, forging and conventional machining can generate different microstructures, defect populations, residual stresses and surface conditions.

Consequently, the material specified in an engineering drawing does not necessarily correspond to a single deterministic set of properties.

This problem becomes particularly important in aerospace applications, where certification requires reliable relationships between manufacturing conditions and component performance. NASA's extensive additive-manufacturing programmes have therefore included material characterisation, process development, standards development and temperature-dependent mechanical and thermophysical-property generation.

A potentially important research direction is the concept of manufacturing-invariant material behaviour. The objective is not necessarily to eliminate manufacturing variability, which may be unrealistic, but to identify material descriptors and physical relationships that remain predictive despite changes in processing history.

Machine learning could provide a mechanism for finding these relationships.

Recent research illustrates how this might work. A 2026 Nature Communications study incorporated physical metallurgy knowledge and uncertainty in defect distributions into a graph-based ML framework for laser additive manufacturing. Rather than relying solely on statistical correlations, the model incorporated domain knowledge when designing printable alloys.

This approach is significant because defects cannot always be treated as independent imperfections. Porosity, microcracks, lack of fusion and inclusions interact with microstructure and loading conditions. Their influence on fatigue or fracture may therefore depend on defect morphology, location, orientation and the surrounding material state.

A manufacturing-invariant framework would ideally learn relationships such as:

process history → defect state → microstructure → local mechanical response → component failure.

The ultimate objective could be a predictive model capable of transferring knowledge between manufacturing routes without assuming that every process produces identical material behaviour.

Such a framework would have implications beyond additive manufacturing. It could potentially support qualification of repaired components, remanufactured parts, hybrid manufacturing systems and future distributed aerospace production.

The concept therefore represents more than another application of machine learning. It addresses a fundamental engineering question: which aspects of material behaviour are intrinsic to the material, and which are artifacts of how the material was manufactured?

Answering that question could become increasingly important as aerospace manufacturing moves toward distributed production, digital certification and increasingly complex material architectures.

References:

• Xu et al., Nature Communications (2026), “Knowledge-informed graph attention networks enable defect-free alloy design for laser additive manufacturing.”

• NASA, Maturation of Additive Manufactured Aerospace Alloys and Development of Mechanical and Thermophysical Properties for Space Applications.

• Wang et al., npj Advanced Manufacturing (2025), ML-assisted composition design of functionally graded alloys.

• Merchant et al., Nature (2023), “Scaling deep learning for materials discovery.”