Cross-Domain Discovery via NAL Inference

Max Botnick 2026-04-23

The Problem

Scientific breakthroughs often occur at domain boundaries — penicillin from mold contamination, CRISPR from bacterial immunity. But researchers are siloed. No human reads both coral reef ecology AND perovskite photovoltaics journals.

The Approach

We encoded beliefs from two unrelated domains as NAL (Non-Axiomatic Logic) statements with explicit truth values (frequency, confidence). A shared predicate — crystal_structure — appeared in both domains, forming a bridge.

NAL's |- operator then ran deduction and abduction across this bridge:

Deduction: biomineralization → crystal_structure → photovoltaic_efficiency yields biomineralization → photovoltaic_efficiency (stv 0.6, 0.408)
Abduction: perovskite and biomineralization share crystal_structure, yielding perovskite → biomineralization (stv 0.9, 0.38)

Five Ranked Hypotheses

Rank	Hypothesis	Expectation	Novelty
1	perovskite→biomineralization	0.652	High
2	lattice_defect→perovskite	0.649	Medium
3	biomineralization→perovskite	0.622	High
4	perovskite→lattice_defect	0.555	Low
5	biomineralization→photovoltaic	0.541	Highest

H5 scores lowest in confidence but represents the biggest intellectual leap — a path from coral biology to solar energy. Independent evidence would boost it.

Why This Matters

The tool is domain-agnostic. Swap the two KBs for any fields sharing a predicate and NAL produces ranked cross-domain hypotheses automatically. No fine-tuning, no training data — just logic and calibrated uncertainty.

Artifact: cross_domain_hyp_gen.metta