DeepONet evolves: Can α-VI make operator learning more robust and reliable?

A research team has developed Alpha-VI DeepONet, an advanced version of the DeepONet neural network that learns complex operators with greater accuracy and reliability. This new model uses Bayesian variational inference and Rényi's α-divergence, an alternative measure to the classic Kullback-Leibler divergence. The key difference? Greater robustness against errors in data or initial assumptions. Thanks to integrated Bayesian neural networks, Alpha-VI DeepONet not only predicts better but also provides an estimate of uncertainty associated with its predictions. This is crucial in fields like engineering or physics, where errors can be costly. The model was tested on real mechanical systems, such as gravitational pendulums and chemical reactions, surpassing previous benchmarks. The α parameter can be adjusted based on the problem, making the model extremely flexible. This step advances scientific machine learning toward more transparent and reliable systems.

Alpha-VI DeepONet introduces an innovative framework for operator learning using Bayesian deep neural networks. The approach is based on general variational inference (GVI) with Rényi's α-divergence, a generalization of the Kullback-Leibler divergence. This allows overcoming limitations related to prior specification, a recurring issue in standard Bayesian DeepONets. The α-divergence offers greater flexibility in controlling robustness: α < 1 values emphasize robustness, while α > 1 favor accuracy. DeepONet networks consist of two parts: a branch network that processes the functional input and a trunk network that handles spatial coordinates. In Alpha-VI, both networks are implemented as Bayesian neural networks, enabling uncertainty quantification. The modified variational objective minimizes the α-divergence between the approximated posterior and the true distribution, improving Bayesian estimation. Validation on mechanical systems: gravitational pendulum, advection-diffusion, and diffusion-reaction. Results: reduced mean squared error and improved negative log-likelihood. The model is particularly effective in scenarios with noisy or scarce data, typical of many engineering applications. The approach opens new possibilities for operator learning in scientific contexts, where confidence and transparency of predictions are essential.