Neuro-Symbolic Hypothesis Engine: A Unified Architecture for Autonomous Scientific Hypothesis Generation

The exponentially growing body of scientific literature has made manual synthesis and hypothesis generation increasingly impractical, introducing an essential bottleneck in the scientific discovery pipeline. While large language models (LLMs) have unprecedented capability to process and summarize textual knowledge, their outputs tend to lack scientific foundations, outputting statistically correct but logically inconsistent or physically implausible statements. To fill this gap, we present the Neurosymbolic Hypothesis Engine (NSHE), an original, end-to-end trainable architecture combining both neural and symbolic AI paradigms—comprising a large language model (Qwen), a spiking neural network (SNN) for integration of temporal concepts, an energy-based transformer (EBT) for scalable plausibility scoring, and symbolic knowledge retrieval utilities—into one cohesive framework supporting autonomous, traceable, and scientifically legitimate hypothesis generation. We demonstrate NSHE’s effectiveness by applying it to three high-impact materials science areas: next-generation battery materials, photovoltaic efficiency improvement, and nanomaterials for water purification. Beginning from natural language inquiry, NSHE identifies structured scientific concepts, retrieves and contextualizes recent work from arXiv, and outputs hypotheses not just original but specific, testable, and constrained by knowledge in the domains. Importantly, the system outputs an auditable reasoning trace, reducing it from blackbox predictor to cooperative, interpretable collaborator in scientific discovery. This work constitutes an initial step in creating AI systems to actively accelerate scientific innovation while conformed to the epistemological rigor of the scientific method.