Nikolas: Agent for Autonomous Catalyst Discovery Using the PRISM Meta-Cognitive Architecture

The discovery of efficient catalysts for CO 2 reduction is critical for addressing climate change, yet traditional approaches rely on expensive noble metals and timeconsuming trial-and-error experimentation. We present Nikolas, a semi-autonomous AI agent for scientific discovery, powered by PRISM (Persistent Recursive Intelligence with Structured Metacognition)-a novel 5-layer meta-cognitive architecture that enables emergent domain intuition from the interplay of large language models (LLMs), computational tools, and hierarchical memory. In its first application, Nikolas autonomously generates and screens catalyst candidates: it proposes structures using Gemini 2.5 Flash, validates them with RDKit, calculates CO 2 binding energies using xTB (GFN2-xTB), and filters candidates using the Sabatier principle (optimal binding: −10 to −20 kcal/mol). The agent's automated screening loop identified a champion catalyst, Fe(2-pyridine) 2 (PPh 3) 2 , with −15.45 kcal/mol CO 2 binding energy (Sabatier optimal). The researchers then independently validated this candidate through reaction pathway analysis, 50 ps molecular dynamics simulation (zero ligand detachment), and hydrogen evolution reaction (HER) selectivity testing (11.64 kcal/mol thermodynamic advantage over H 2 binding). While this paper demonstrates Nikolas in the domain of computational chemistry, the PRISM architecture is domain-agnostic-designed to accelerate scientific discovery across disciplines from drug design to materials science.