Beyond Semantic Similarity: Contrastive Learning for Cross-Domain Structural Isomorphism Detection (v1.1 Expanded)

v1.2 (April 2026) — V2 Model Strict Screening Results This version adds the complete V2 strict pipeline outputs on top of v1.1's V1 broad pipeline: V2 model trained on 5689 expanded samples (1214 SIBD + 4475 new KB). Silhouette 0.55, Retrieval@5 96% on expanded 4443-phenomenon test set. V2 strict retrieval: 4533 cross-domain high-similarity pairs (threshold 0.70), 75× stricter than V1's 339,913 pairs on the same KB. 50-batch LLM strict screening: 94 five-score pairs (strictly isomorphic), 667 four-score (total 761 / 16.8% high potential). 10-batch deep analysis: 94 top pairs analyzed across 5 dimensions → 19 A-level candidate papers, 22 B+, 36 B, 17 C (downgrade rate shows 5/5 filter is not sufficient). V1 vs V2 comparison: zero overlap between V1's 24 tier-1 and V2's 19 A-level discoveries — the two pipelines capture entirely different structural isomorphism signals and are complementary, not redundant. Top 5 V2 A-level findings: Permafrost methane delayed feedback × Extinction debt (8.6); Semiconductor laser relaxation oscillation × Algorithmic stablecoin (8.6); Percolation threshold × Technology adoption chasm (8.5); MHC over-dominant selection × Model ensemble (8.5); Extinction debt × ENSO delayed oscillator (8.4). V3 pipeline design included as plans/v3-pipeline-design.md — proposes StructTuple structured representation + LLM pairwise rerank, predicted to lift A-level hit rate from 20% to 35-45%. Together with v1.1's V1 pipeline outputs (272 A-rated, 24 tier-1), this release provides 43 unique top-tier cross-domain structural isomorphism candidates spanning computer science, physics, biology, ecology, finance, and engineering. Corresponding V2 model weights: qinghuiwan/structural-isomorphism-v2-expanded on Hugging Face. Version 1.1 - Expanded: Major update with 9x larger knowledge base. We present a contrastive learning approach to detect deep structural isomorphism across scientific domains, moving beyond surface semantic similarity. Key results from the expanded v1.1: SIBD Expanded Benchmark: From 500 to 4,443 scientific phenomena (8.9x growth) Large-scale Discovery: 9,867,903 pairwise comparisons yielding 339,913 high-similarity cross-domain matches (112x more than v1) Rigorous Screening: 5,000 top matches screened, 3,167 passed (4.6x more), 272 A-rated, 24 tier-1 discoveries (4x more), 5 score-9 top findings (2.5x more) Fine-tuned Model: Silhouette 0.85, Retrieval@5 100% Key Tier-1 Findings: Quantum Error Mitigation × Quantization-Aware Training (score 9) Path Integral Duality × Neural Tangent Kernel (score 8) Signaling Game × RLHF Alignment (score 8) Mast Cell Degranulation × DeFi Liquidation Cascade (score 8) All code, data, and models are open-source at GitHub.