BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:Physics-Informed Deep Learning and Bayesian Inference for Inverse 
 Problems in 1D ¹H NMR Analysis
DTSTART:20261028T140000Z
DTEND:20261028T150000Z
DTSTAMP:20261103T055000Z
UID:indico-event-19504@indico.psi.ch
CONTACT:matthias.krack@psi.ch
DESCRIPTION:Speakers: Nicolas Schmid (ZHAW)\n\nNuclear Magnetic Resonance 
 (NMR) spectroscopy probes local magnetic environments to characterize mole
 cular structures\, serving as a standard experimental tool across chemistr
 y and materials research. However\, extracting underlying physical paramet
 ers from experimental 1D ¹H spectra behaves as an ill-posed inverse probl
 em. Signal overlap\, experimental artefacts\, and quantum-mechanical stron
 g coupling pose persistent challenges. When scalar couplings approach the 
 magnitude of chemical-shift differences\, off-diagonal Hamiltonian terms m
 ix spin states\, shifting transition frequencies and redistributing intens
 ities in ways that cannot be determined by simple linear first-order rules
 .\nMuch like forward-modelling workflows in condensed-matter physics and e
 lectronic structure theory\, forward simulations from a spin Hamiltonian c
 an be carried out reliably\, but the corresponding inverse mapping tends t
 o be non-unique and numerically challenging. We present a hierarchical fra
 mework that addresses these difficulties by combining physics-informed dee
 p learning\, quantum spin dynamics\, and Bayesian inference across three c
 omplementary stages:\n\nPhysics-Informed Deconvolution (MLdecon): Rather t
 han fitting unconstrained peak profiles\, a convolutional neural network t
 rained on synthetic spectra generated from physical line models learns to 
 decompose congested envelopes into peak positions\, widths\, and intensiti
 es. In benchmark tests\, this model shows competitive performance against 
 standard optimization routines and has been integrated into Bruker TopSpin
  for laboratory workflows.\nDirect Hamiltonian Parameter Detection (MolDET
 R): To move beyond local peak fitting\, we adapted a set-based detection t
 ransformer to estimate spin-system descriptors directly from 1D ¹H spectr
 a. Trained on synthetic data generated through quantum-mechanical spin-dyn
 amics simulations with artefact augmentation\, MolDETR predicts unordered 
 sets of chemical shifts\, scalar coupling constants\, and proton counts wi
 thout requiring a known chemical graph. On experimental test sets\, it ach
 ieves median errors typically below 1 Hz for chemical shifts and around 0.
 2 Hz for scalar couplings.\nBayesian Refinement with Explicit Quantum Hami
 ltonians: Empirical priors (such as HOSE codes or standard DFT estimates) 
 can be broad and occasionally misguide non-linear solvers. Using MolDETR p
 redictions as concentrated\, spectrum-conditioned priors helps constrain B
 ayesian total line-shape analysis under an explicit quantum spin Hamiltoni
 an. Sequential Monte Carlo and Hamiltonian Monte Carlo sampling explore th
 e resulting parameter spaces\, offering posterior distributions with physi
 cally calibrated uncertainty estimates.\n\nBy coupling synthetic spin simu
 lations with transformer-based inference and Bayesian sampling\, this work
 flow highlights a practical strategy for handling complex inverse problems
  and parameter estimation in experimental spectroscopy.\n\nhttps://indico.
 psi.ch/event/19504/
LOCATION:OVGA/200
URL:https://indico.psi.ch/event/19504/
END:VEVENT
END:VCALENDAR