Article View

An Autonomous Scientific Discovery Platform Using AI Agents, Knowledge Graphs, and Reproducible Data Pipelines

DOI: 10.66838/ma.234

Chaitanya Nagabhyru 12-1006, Rahmath Nagar 2nd Line, Chilakaluripet, Guntur Pin: 522616, Andhra Pradesh, India

Abstract

The volume of scientific literature and data is growing faster than any researcher can read or connect, and the same growth makes irreproducible results costly. Large language model agents can search, summarise and write code, but they invent citations, state unsupported claims with confidence and leave no auditable trail. This paper presents a platform for AI-assisted scientific discovery that addresses these weaknesses by design. A knowledge graph, built from literature and structured datasets with evidence spans and provenance attached to every edge, serves as the shared memory of a team of specialised agents. The agents retrieve evidence, propose testable hypotheses, rank them by a transparent score that combines plausibility, evidence support, novelty and expected information gain, and run computational experiments inside containerised, version-controlled pipelines whose every step is logged. An independent verification layer, separate from the generating agents, checks each claim against its cited source, recomputes reported numbers, re-executes the pipeline, and flags unsupported statements. Humans approve every consequential step. The paper specifies the architecture, the formal definitions of the main scores and metrics, and a pre-specified evaluation protocol with four measures: reproducibility, hypothesis-ranking quality, citation accuracy and researcher time saved. The protocol uses time-sliced historical evaluation, in which hypotheses are ranked from a graph frozen at a past date and tested against discoveries made afterwards, together with a controlled within-subject user study. This manuscript presents the design, formal definitions and protocol. It does not report experimental results: the result tables in Section VI are templates to be completed from actual runs, and no performance claim is made.

Keywords

AI for Science, Knowledge Graphs, AI Agents, Reproducibility, Hypothesis Generation, Provenance, Retrieval-Augmented Generation,