A collective of senior researchers working on inverse problems through two neuro-symbolic methods — classical scientific machine learning, and LLM-orchestrated RAG-mediated calibration of complex simulators.
The Institute reads the inverse problem as a precise operational act. In every instance of work, two functional roles are distinguished and brought into relation: a material to be inverted — the heterogeneous evidence on which the conversion operates — and a functional receiver — the structured model that accepts the inverted configuration and uses it. The two roles are not properties of the entities involved; they are states that materialise only at the moment of the inversion.
The two functional roles are not properties of the entities involved. They are states that the entities take on for the duration of a single inversion act. A new instance — a new lake, a new regulatory question, a new time interval — requires the roles to be re-activated.
Measurements, structural priors, and accumulated knowledge of the system are converted into parameter values and closure conditions. The receiver is the deterministic model — ODE compartmental, PDE distributed-parameter, state-space — augmented where needed by PINNs, Universal Differential Equations, or graph neural networks.
A hybrid syntactic-vector RAG indexing a domain corpus, queried under LLM orchestration, is the material from which the inverted configuration is extracted. The receiver is the same category of mechanistic model as in Line B — a simulator that accepts parametric profiles, initial conditions, and boundary structures the RAG-LLM has assembled.
Six sub-areas in total. Line A is articulated in a methodological foundation plus two applied projects on Italian sub-alpine lakes. Line B is articulated in three reformulations of foundational papers from the Institute's lineage.
Targeted Retrieval-Augmented Generation — hybrid syntactic-vector RAG on the EMA EPAR + Product Information corpus. Mature operational prototype.
RAG on Manca's 30+ years of zooplankton data generates the egg age-structures required by a compartmental ODE population model for parametric inversion.
RAG on Lake Varese documentation generates the parametric profiles required to calibrate the General Lake Model + AED2 biogeochemistry module.
McKendrick-Von Foerster age-transport + PINN-based birth-rate inversion + deep image classifiers on plankton microscopy.
NUMSAS lineage (Brussels 1979) and Loss-Pattern Identification (INMM 1982 with LANL), reformulated as state-space neural inversion.
Anatomically-informed graph neural network for compartmental tracer kinetics. Foundation: Merlini 1971 on Lepomis gibbosus.