Swedish TRUSTAM project receives Vinnova funding to bring federated learning AI to additive manufacturing quality control

The Swedish Innovation Agency Vinnova has awarded funding to the TRUSTAM (Trusted Federated Intelligence for Additive Manufacturing) project, a consortium including Interspectral, Saab, AMEXCI, and Scaleout Systems. The project aims to apply federated learning (FL) — a machine learning architecture where AI models train across distributed sites without sharing raw data — to industrial additive manufacturing quality control. Interspectral will lead development of local AI models that learn from each machine's unique process data, integrating monitoring, analysis, and decision-making into a continuous operational loop. The project runs through early 2028, culminating in full demonstrations and dissemination in aerospace and defense environments where quality traceability is non-negotiable.

This partnership directly addresses a structural bottleneck that has constrained AI adoption in safety-critical AM: the tension between collective intelligence and data sovereignty. As metal AM moves into regulated production paths in defense, aerospace, maritime, and energy, data management and cybersecurity have become primary scaling barriers, not secondary concerns. Interspectral's existing AM Explorer platform, already integrated into over 60% of metal AM machines and deployed at customers including GKN Aerospace and Volum-E, provides a mature foundation. The federated learning approach distinguishes TRUSTAM from single-site AI quality tools like Ai Build's path generation or Oqton's build quality platform, which improve first-print success rates and combine simulation, monitoring, and inspection but cannot share intelligence across sites without exposing proprietary process data.

For Interspectral, this project validates its platform strategy and positions it to own the infrastructure layer for distributed, secure AI in AM quality assurance. The practical challenge is execution: federated learning in heterogeneous machine fleets with varying sensor configurations and process signatures is technically demanding, and the aerospace and defense qualification grind means any output must be verifiable, not just intelligent. If the consortium delivers a working framework by 2028, it will have built something the industry currently lacks — a way to pool learning across production sites without compromising the data sovereignty that primes, primes, primes.