Methods for identification of latent structures from observed data.
A forward problem runs a known mechanism to predict what will be observed. An inverse problem runs the other way: from what was observed, back to the mechanism that produced it. The Institute exists for the second direction.
The inversion act
Nature, instruments and archives give us traces: population counts, signals, balances, documents. What produced them — the rates, the parameters, the hidden states — is never observed directly. Recovering it is inference under constraint, and it is the same act whether the substrate is a lake, a nuclear facility or a regulatory corpus.
Hence the Institute's claim: the method is the object. Domains change, the inversion does not.
When the know how was developed
Four decades across three European Commission divisions:
- JRC Ispra, AI Laboratory — symbolic reasoning, analogical representation, computational linguistics.
- Computational nuclear safeguards — statistical material accountancy and loss-pattern detection under strict error control.
- European Medicines Agency, London — regulatory informatics over large documentary corpora; the direct antecedent of Line A.
Inversion through RAG-LLM
A large language model coupled to retrieval over a domain corpus converts accumulated documentary knowledge into the quantities a simulator needs. Retrieval and generation become instruments of inference, not a search convenience.
Foundation: TRAG — hybrid syntactic-vector retrieval over ~3,200 regulatory documents, running as a mature prototype.
Scientific machine learning
Differential-equation models of a mechanism, with neural components carrying what resists closed form. Estimation stays under symbolic constraint: the network is bounded by the physics, not substituted for it.
Lineage: continuous from the 1973 compartmental work at Ispra to the demographic estimators under review in 2026.