Chenyang Zhu

Fluid migration screening requires the integration of seismic structure and well log response. This study develops a multimodal framework for ranking an attribute-derived fluid migration favorability proxy from paired local seismic patches and well log windows in the F3 survey. The framework couples frozen seismic foundation model (SFM) features with a trainable well log token encoder and evaluates pooled concatenation, cross attention, and learned query fusion for multimodal integration. Under sample-stratified evaluation across three random seeds, frozen SFM features with learned query fusion achieved the highest observed normalized discounted cumulative gain at rank 20 (nDCG@20) of 85.15%, exceeding the best single-modality baseline by 3.95 percentage points. The results support combining pretrained seismic representations with well log evidence to prioritize intervals in the F3 benchmark.

Reservoir sweet spot screening combines seismic and well-log observations, but an observed gain may depend on the input representation, validation protocol, and construction of the screening target. We implement a controlled performance-decomposition protocol for paired observations from the public F3 Block. The protocol formalizes matched nested comparisons among log-only, basic seismic–log, and directional-gradient seismic–log representations. Within each contrast, the outer test observations, metric, model family, and validation-only selection rule remain fixed. For the composite target, basic fusion increases mean area under the receiver operating characteristic curve (AUROC) by 6.5, 8.8, and 3.4 percentage points for logistic regression, random forest, and gradient boosting, respectively. Directional-gradient enrichment changes AUROC by −5.6, +1.5, and −2.8 percentage points in the same model order. Spatial-block AUROC ranges from 49.4% to 57.8%, while leave-one-well-out AUROC ranges from 42.9% to 53.9%. The corresponding basic-fusion contrasts are −8.5 and +2.2 percentage points. A target-overlap audit shows that the measured fusion gain depends on target construction and directly corresponding log channels. Together, these controlled comparisons quantify how representation design, target construction, and validation protocol shape reported reservoir-screening performance.

Reinforcement learning (RL) in partially observable and noisy environments remains a central challenge because of the difficulty in inferring latent structures, managing label uncertainty, and ensuring robust policy optimization. Reward Machines (RMs) offer a structured formalism for representing non-Markovian rewards through symbolic automata, but existing methods exhibit fragility under stochastic observations. To address this limitation, we introduce Probabilistic Induction with Genetic Local Search (PI-GLS), a unified framework that integrates probabilistic perception, inductive logic programming, and evolutionary optimization. PI-GLS achieves noise-robust RM induction by combining Bayesian inference for uncertain labels, sampling-based symbolic abstraction, and genetic local refinement of automaton structures. In addition, we propose a belief-aware reward shaping strategy that leverages distributions over RM states to guide policy learning under uncertainty. Extensive experiments on benchmark domains show that PI-GLS substantially improves convergence efficiency, robustness to sensor noise, and interpretability of learned reward models, achieving performance on par with manually engineered RMs even under severe noise. These results demonstrate the scalability and effectiveness of PI-GLS for autonomous decision-making in real-world scenarios where stochastic feedback and partial observability are unavoidable.

Safety-critical decision-making often necessitates operating within complex and uncertain environments, typically characterized as partially-observable Multi-Agent Systems (MAS). Effective decision-making in these settings demands a profound understanding of the environment and accurate representation and quantification of epistemic uncertainty due to informational deficits. This uncertainty representation must reflect prior knowledge and dynamically integrate observational data. While Version Space Learning (VSL) offers a framework for set-based uncertainty representation, prior methods often depend on under-approximation of the actual version space, which falls short in safety-critical applications. We introduce a model checking-based VSL framework to enhance uncertainty representation and context-awareness in such applications. Our approach employs network of timed automata (NTA) to model the MAS environment, capturing its mechanisms and prior knowledge efficiently. The inherent non-determinism of timed automata facilitates an over-approximation of the actual version space, ensuring all plausible hypotheses are considered. We further refine the parameter ranges of the NTA using proof traces, optimizing the use of observational data. In a medical diagnosis case study, our framework identified a missing rule in a traditional rule-based system and provided interpretable and scalable results, demonstrating our method’s ability to improve decision-making accuracy and manage complex scenarios effectively.