Přístupnostní navigace
E-application
Search Search Close
Master's Thesis
Author of thesis: Ing. David Jedlička
Acad. year: 2025/2026
Supervisor: doc. Mgr. Petr Vašík, Ph.D.
Reviewer: Ing. Roman Byrtus
In this thesis, we focus on the design of an equivariant neural network based on Clifford algebras for the autoregressive prediction of physical field evolution. The contribution of this work is the design and implementation of three autoregressive architectures for the prediction of the nonhomogeneous heat equation. The primary model, CGELSTMFNO, combines a Clifford group equivariant encoder and decoder, an equivariant recurrent module CGELSTM that extends the classical LSTM architecture into the framework of Clifford algebra, and a stack of CGEFourier blocks that perform spectral convolution. Two additional architectures, HybridLSTMFNO and ClassicLSTMFNO, serve as a comparison with CGELSTMFNO. Experiments were conducted across three diffusion regimes, and the models were evaluated on test horizons that were twice as long as the training sequences. The results show that both Clifford-based models consistently outperform the classical model despite using a comparable number of parameters and the same training data. The comparison between CGELSTMFNO and HybridLSTMFNO suggests that the primary benefit of Clifford algebra in the proposed models arises from equivariant processing of spatial information rather than from the recurrent module itself.
Clifford algebra, Clifford group, Equivariant neural networks, Recurrent neural networks, Neural operators, Heat equation
Date of defence
08.06.2026
Result of the defence
Defended (thesis was successfully defended)
Grading
A
Process of defence
Student prezentoval práci, školitel i oponent přečetli své posudky. K otázkám oponenta: Zkoušel autor jiný počet bloků FNO? Zkoušel, ale nemělo to moc vliv, tak zůstal u použitého počtu. Jaké aktivační klíče byly využity? Různé v závislosti na modelu, vše vysvětleno (tanh, sigmoid, GeLU...). doc.Vašík: závislost na koeficientu difuze, určení vhodné hodnoty alfa dr.Eryganov: Proč se modely špatně chovají na okrajích? Vše zodpovězeno.
Language of thesis
English
Faculty
Fakulta strojního inženýrství
Department
Institute of Mathematics
Study programme
Mathematical Engineering (N-MAI-P)
Composition of Committee
doc. Mgr. Petr Vašík, Ph.D. (předseda) doc. RNDr. Martin Kolář, Ph.D. (místopředseda) prof. Aleksandre Lomtatidze, DrSc. (člen) Ing. Ivan Eryganov, Ph.D. (člen) Ing. Petra Rozehnalová, Ph.D. (člen)
Supervisor’s reportdoc. Mgr. Petr Vašík, Ph.D.
Grade proposed by supervisor: A
Reviewer’s reportIng. Roman Byrtus
Grade proposed by reviewer: B
Responsibility: Mgr. et Mgr. Hana Odstrčilová