TL;DR
Fed-PIGANO is a new federated learning framework that allows distributed manufacturing sites to collaboratively train thermal prediction models without sharing their proprietary part designs. By combining physics-informed neural operators with federated aggregation, the system achieves high accuracy on unseen geometries, performing nearly as well as a centralized model while maintaining strict data privacy.
What Problem Does This Paper Address?
In metal additive manufacturing (AM), the thermal history of a part—how it heats up and cools down during the build—is the "holy grail" for predicting final quality, microstructure, and mechanical properties. However, calculating this thermal field using traditional numerical methods like the Finite Element Method (FEM) is computationally expensive, especially when designers need to iterate through hundreds of potential part shapes.
While "geometry-aware" machine learning models can predict these fields in milliseconds, they require a massive, diverse dataset of part geometries to learn effectively. In the real world, this data is siloed. Manufacturing sites treat their part designs as highly sensitive intellectual property (IP). Contractual or regulatory constraints often make pooling these designs into a single central database impossible. As a result, a model trained only on one site’s limited data generalizes poorly to new, complex shapes.
Methods
The authors propose Fed-PIGANO (Federated Physics-Informed Geometry-Aware Neural Operator). The framework consists of two main pillars:
- Local Geometry-Aware Modeling (PI-GANO): Each site uses a neural operator architecture that takes a part's boundary coordinates as input. It uses a geometry encoder (similar to PointNet) to create a "shape embedding" and a trunk/decoder network to predict the temperature at any coordinate within that shape. Crucially, it is physics-informed: the model is trained to minimize the residual of the heat conduction partial differential equation (PDE) and boundary conditions, meaning it doesn't necessarily need expensive "labeled" simulation data to learn the underlying physics.
- Federated Learning (FedAvg-Adam): Instead of sharing geometries, sites share only their model weights. A central server aggregates these weights to create a global model and broadcasts it back. To handle the tricky optimization landscape of physics-informed models, the authors use a modified strategy where Adam optimizer moments are maintained locally at each site across communication rounds, stabilizing the learning process.
Datasets
The researchers evaluated the framework using a 2D steady-state thermal benchmark representing a laser spot-melting process.
- Geometry Families: Five distinct polygonal families were used, ranging from convex (Quadrilaterals, Pentagons, Hexagons) to concave (Concave-Pentagons, Concave-Hexagons).
- Distribution Scenarios: They tested four scenarios (S1–S4) with increasing levels of "heterogeneity." In S1, all sites had a balanced mix of all shapes. In S4, sites had highly specialized datasets (e.g., Site 1 only saw convex shapes, while Site 3 only saw concave ones).
- Scale: Each site held a training set of 50 geometries, with 300 geometries per family total across the network.
Key Insights and Ideas
- Federation as a Diversity Multiplier: The primary benefit of federation here isn't just privacy—it's geometric coverage. A site that only manufactures simple brackets can suddenly predict thermal fields for complex turbine blades because it has benefited from the "physics knowledge" encoded in the weights of a model trained elsewhere.
- Physics as the Common Language: Because every site is solving the same underlying heat equation, the model parameters are inherently compatible. The physics acts as a regularizer that makes federated aggregation meaningful even when the training geometries look nothing like each other.
- Stable PINN Federation: The paper highlights that standard federated averaging (FedAvg) can struggle with the complex loss functions of physics-informed networks. Keeping the Adam optimizer's state local to each client is a key technical detail that ensures the global model actually converges.
Key Results
- Accuracy: Fed-PIGANO achieved a mean relative $L_2$ error of 1.83%–2.04% across all test cases.
- Superiority over Local Training: In the most heterogeneous scenario (S4), a site training only on its own data had errors as high as 20.46% when faced with unseen geometry types. Fed-PIGANO reduced this error by an order of magnitude.
- Parity with Centralized Models: Remarkably, the federated model performed almost identically to a "Centralized" reference model that had full access to all data in one place—a setup that is usually impossible in practice due to IP concerns.
Limitations and Future Work
- Dimensionality and Time: The current study is limited to 2D steady-state problems. Real-world AM is 3D and transient (the heat source moves over time). Extending this to 3D and time-dependent PDEs is the next logical step.
- Coupled Physics: The authors suggest investigating how this framework could handle multi-physics problems, such as coupling thermal fields with microstructure evolution models.
- Enhanced Privacy: While geometries aren't shared, advanced "model inversion" attacks can sometimes reconstruct training data from weights. The authors note that future versions could incorporate differential privacy or secure multi-party computation to provide even stronger IP protection.
Summary updated