Introduction
In fields such as materials science, researchers often face the challenge of sparse and heterogeneous data. High-quality experimental results are expensive to obtain, forcing scientists to rely on a "fusion" of disparate sources, including legacy patents, journal articles, and limited in-house experiments. However, these sources are rarely consistent; they may involve different experimental setups, measurement errors, or unreported "hidden" physical parameters. Standard machine learning models typically treat all data points as originating from a single distribution, often leading to high uncertainty and poor predictive accuracy when sources conflict.
This paper introduces a "source-aware" data fusion framework based on Latent Variable Gaussian Processes (LVGP). By mapping categorical data sources into a physically interpretable latent space, the framework improves prediction accuracy and provides a mathematical basis for quantifying source dissimilarity.
The LVGP Framework
The core innovation of this approach is the use of LVGP to handle categorical variables (e.g., "Source A" vs. "Source B") for which no natural distance metric exists. The model functions through a multi-step process:
- Latent Mapping: Each data source is treated as a categorical variable and mapped into a low-dimensional (typically 2D) continuous latent space.
- Statistical Inference: During training, the model learns the optimal coordinates for each source in this space. Proximity in the latent space indicates similarity in the underlying relationship between inputs and outputs.
- Dissimilarity Metric ($D$): A normalized Euclidean distance in the latent space allows researchers to rank external sources against a trusted "Ground Source."
- Source-Specific Prediction: The model generates smoother response surfaces by accounting for the specific "coordinates" of each source, effectively filtering noise from lower-quality data.
Experimental Validation
The framework was validated across several case studies, ranging from synthetic benchmarks to real-world alloy datasets.
Synthetic Parabola Benchmark
To test the model's fundamental capabilities, the authors used a synthetic problem involving four different equations: one "Ground Source" and three "Perturbed Sources." The LVGP achieved a Testing Normalized Root Mean Squared Error (NRMSE) of $11.66 \times 10^{-4}$, a dramatic improvement over the $0.207$ NRMSE produced by a standard source-unaware Gaussian Process.
Materials Science Applications
- FeCrAl Alloys: In a study of thermal aging embrittlement involving 380 data points from GE Research and seven literature sources, the source-aware LVGP reduced error by approximately 25% compared to a standard GP.
- SmCoFe Alloys: Fusing data from two patents and one peer-reviewed paper (71 points total) to predict magnetic energy products resulted in uncertainty levels three times lower than a single-source model.
Key Insights: The Latent Space as a Physics Proxy
A significant finding is that the learned latent space often reflects physical reality. In the FeCrAl study, sources with similar experimental procedures clustered together, while outliers corresponded to known differences in manufacturing processes. This interpretability allows for "targeted source selection." By using the dissimilarity metric to prune the most distant sources, the authors improved prediction accuracy on target data by an additional 20%. This demonstrates that while "more data" is often sought, "better-aligned data" is more valuable.
Limitations and Future Directions
While powerful, the LVGP framework currently faces scalability challenges, struggling when the number of sources exceeds 15 due to the increasing number of latent variables. Additionally, the current model assumes all sources share the same input parameters. Future work will focus on extending the model to "heterogeneous" cases where sources report different variables and implementing Sparse-Variational LVGP to handle larger datasets.
Summary updated