Projects

Two lines, six sub-areas.

Every project is the same act on a different substrate: recovering a mechanism from the traces it leaves. Line A pursues it through retrieval and language models; Line B through differential equations with neural components.

Line A — frontier

Methodological foundation

TRAG — Targeted Retrieval-Augmented Generation

Hybrid syntactic-vector retrieval and LLM generation over a dataset of several thousand EU regulatory documents. In evolution as a mature prototype. The instrument from which both applied projects derive.

A·1 · applied · partner: Marina Manca

Neuro-symbolic model identification for ecosystem studies and management

Long-term Lake Maggiore zooplankton observations are used to derive and evaluate models.

A·2 · applied

RAG-driven calibration of ecosystem simulation models

Documents and data from Lake Varese will be collected to calibrate a number of lake simulation models.

Methodological collaboration. Leverhulme Centre for Wildfires (KCL + Imperial), contact Oliver Perkins. Simulator, retrieval and language-model fusion. The Institute does not pursue the wildfire domain itself.

Line B — consolidated

B·1 · Memorie 45, 1987 → J. Limnology, 2026

Inverse problems for ecosystem studies

Population dynamics as a tool for understanding the role of zooplankton in the stability of aquatic ecosystems.

B·2 · NUMSAS 1979 → Vienna 1982 → PROSA 1987

Material accountancy & pattern recognition

State-space neural networks and normalising flows for loss-pattern detection in near-real-time nuclear material accountancy.

B·3 · Merlini, 1971

Inverse physiology & radioecology

Anatomical graph neural networks over compartmental kinetics, reopening the Zn-65 physiology of Lepomis gibbosus.

The common thread. A mechanism written down explicitly, a neural component carrying what resists closed form, an estimate delivered with its uncertainty.  Five decades of source material →