Reproduce “Neural Message Passing for Objective-Based Uncertainty Quantification and Optimal Experimental Design” on OpenResearch

About this paper

Qihua Chen, Xuejin Chen, Hyun-Myung Woo, Byung-Jun Yoon

Researchers introduced a neural message-passing network (MPNN) as a data-driven surrogate model to directly estimate the Mean Objective Cost of Uncertainty (MOCU), overcoming the high computational expense of traditional MOCU estimation in Optimal Experimental Design (OED). This approach, enhanced by a novel axiomatic constraint loss, achieved MOCU prediction accuracy with mean squared errors as low as 1.534e-4 and delivered computational speed-ups of four to five orders of magnitude in OED for uncertain Kuramoto models.

Key insights

• Deep learning surrogate models, specifically MPNNs, can accurately and efficiently estimate MOCU, achieving a significant acceleration over traditional sampling-based methods without compromising OED performance.
• Incorporating domain-specific axiomatic constraints into the training loss function enhances model stability and predictive accuracy, particularly in preserving the crucial relative ranking of experiments for OED.
• The computational burden of MOCU-based OED can be shifted from expensive online calculations to an offline training phase, making this principled uncertainty quantification framework more accessible for complex systems.

Summary by alphaXiv.