Sondipon Adhikari · Glasgow

Research theme five of five

Computational mechanics and scientific computing

Exact results still earn their place. They are the only way to know whether an approximation, including a learned one, is telling the truth.

An exact solution and a discretisation of it. Benchmarks with known answers are what make claims about accuracy checkable.

The problem

Scientific machine learning is entering structural dynamics faster than the error analysis that should accompany it. A surrogate that reproduces a training set to three significant figures can still be wrong in exactly the region a reliability calculation depends on, because failure probabilities are governed by rare events, and a training set spreads its points over the bulk.

The position taken here is to work on what the network is asked to learn. It is to supply the layer underneath: exact solutions, reference problems with published answers, and error statements that a regulator could apply.

Why it matters

An approximate method is trusted on the strength of the exact case it was checked against, so exact and semi-analytical results are the instruments that calibrate everything else. Dynamic stiffness carries this furthest: a frequency-dependent element description gives the exact response of a member at any frequency, so a structure that a mesh would need thousands of elements to resolve at high frequency is handled by a handful. The same discipline governs the machine-learning work here. A surrogate is judged by the decision it supports, so the question is where to place training points so that a failure probability moves, and a reproducible benchmark is what lets anyone else check the claim.

Current frontier

  • exact and semi-analytical mechanics
  • dynamic stiffness and exact member formulations
  • spectral and special-function solutions
  • exact stochastic operators
  • stochastic computation and reduced-order modelling
  • neural operators, graph networks and measure operators
  • physics-informed and physics-encoded methods
  • open benchmarks and reproducible examples

Signature concepts

  • exact benchmarks
  • dynamic stiffness
  • model reduction
  • operator learning
  • measure-based computation

Open problems

  • What error bound can be placed on a failure probability computed through a trained surrogate?
  • Where should training points be placed so that they change a reliability estimate?
  • Which exact stochastic results can serve as reference solutions for operator learning?

Selected papers

  • Exact dynamic stiffness formulations and vibration response analysis of orthotropic viscoelastic plate built-up structures
    X. Liu, X. Liu, S. Adhikari · Computers & Structures 302, 107455 · 2024
  • MATLAB implementation of physics informed deep neural networks for forward and inverse structural vibration problems
    T. Chatterjee, M. I. Friswell, S. Adhikari, H. H. Khodaparast · Aerospace Research Communications 2 · 2024
  • Exact energy harvesting analysis of multimodal piezoelectric beams using the dynamic stiffness method
    X. Liu, Y. Wang, S. Adhikari, W. Zhou · Computers & Structures 313, 107746 · 2025
  • A partitioned combined computational method for multi-scale dynamic systems
    P. Yuan, S. Adhikari, Y. Dong · International Journal for Numerical Methods in Engineering 124(16) · 2023
  • Bridging proper orthogonal decomposition methods and augmented Newton-Krylov algorithms
    S. Adhikari and co-workers · 2011

Read more papers in this area

Foundations

Finite element analysis of damped systems, dynamic stiffness formulations, modal and reduced-order methods, and Monte Carlo and stochastic reduced models.