TLDR: A new research paper introduces a generalized data-driven learning method for history-dependent materials, capable of integrating both high-fidelity (costly, accurate) and low-fidelity (cheap, noisy) data. It uniquely quantifies and separates model error (epistemic uncertainty) from inherent data noise (aleatoric uncertainty) using Variance estimation Bayesian recurrent neural networks (VeBRNNs). This enhances prediction reliability, especially in data-scarce scenarios, and offers a robust framework for future applications in material design and analysis under uncertainty.
Understanding how materials behave under different conditions, especially over time, is crucial for engineering and scientific advancements. Traditionally, this involves extensive and costly experiments or simulations to gather high-quality, or ‘high-fidelity,’ data. However, generating large amounts of this premium data can be incredibly challenging due to time and resource constraints.
Conversely, ‘low-fidelity’ data is much quicker and cheaper to produce, but it comes with higher errors and often contains more inherent noise. This creates a dilemma: how can we leverage both types of data effectively to build reliable models?
A recent research paper introduces a groundbreaking approach that generalizes data-driven learning to tackle this very challenge. The method, detailed in the paper “Single- to Multi-Fidelity History-Dependent Learning with Uncertainty Quantification and Disentanglement: Application to Data-Driven Constitutive Modeling”, allows for learning from diverse data sources, including those that change over time (history-dependent) and those that are noisy.
One of the most significant contributions of this work is its ability to quantify and separate different types of uncertainties. In machine learning, there are two main kinds: ‘epistemic uncertainty,’ which represents the model’s own error and can be reduced with more data, and ‘aleatoric uncertainty,’ which is the inherent noise in the data itself and cannot be eliminated, even with infinite data. By disentangling these, the models become more trustworthy, providing clearer insights into where the predictions might be less certain due to model limitations versus natural data variations.
The core of this new methodology lies in its flexible and hierarchical structure, adapting to various learning scenarios. It builds upon advanced neural networks, specifically a type called Variance estimation Bayesian recurrent neural networks (VeBRNNs). These networks are particularly well-suited for history-dependent problems, like predicting how materials deform and recover over time.
The training of these VeBRNNs is a cooperative, three-step process. First, a basic network learns the average response. Second, a separate network learns to estimate the data noise (aleatoric uncertainty). Finally, a Bayesian network refines the average prediction and quantifies the model’s own uncertainty (epistemic uncertainty). This modular design means that simpler versions of the model can be used if certain types of uncertainty quantification aren’t needed, making it highly versatile.
For situations involving both high- and low-fidelity data, the researchers propose a multi-fidelity framework. This framework intelligently combines information from different data sources. It starts by training a model on the low-fidelity data, then transfers this learned knowledge to help train a more accurate model using the high-fidelity data. A key finding was that transferring the ‘hidden states’ (internal representations) from the low-fidelity model was more effective than just transferring its direct outputs, suggesting a deeper level of knowledge transfer.
The versatility of this method was demonstrated across four different scenarios, mimicking real-world material modeling challenges. These included cases with only noisy data, combinations of noiseless high-fidelity and noiseless low-fidelity data, and even scenarios where both data types were noisy or where low-fidelity data was noisy but high-fidelity data was clean.
The results showed that multi-fidelity models often outperform single-fidelity models, especially when high-fidelity data is scarce. By strategically using computationally cheaper low-fidelity data, the models could achieve better accuracy and generalize more effectively to new, unseen conditions. The ability to quantify uncertainty also meant that the models could indicate when their predictions were less reliable, particularly for data outside their training range.
Also Read:
- Bridging the Forgetting Divide in AI: A Brain-Inspired Approach to Continual Learning
- AI Model Enhances Soft Body Contact Simulations with Advanced Detection
This research opens up exciting possibilities for future applications in diverse scientific and engineering fields. The ability to learn from mixed-fidelity data while providing clear uncertainty estimates is particularly valuable for designing and analyzing materials under uncertain conditions. The authors envision future work integrating this framework with ‘active learning’ and ‘Bayesian optimization,’ which would allow for intelligent, on-the-fly decisions about which type of data to acquire next for optimal learning or design.
