Details
| Originalsprache | Englisch |
|---|---|
| Seiten (von - bis) | 4569-4588 |
| Seitenumfang | 20 |
| Fachzeitschrift | Engineering with computers |
| Jahrgang | 38 |
| Ausgabenummer | 5 |
| Frühes Online-Datum | 23 Mai 2022 |
| Publikationsstatus | Veröffentlicht - Okt. 2022 |
Abstract
In machine learning, if the training data is independently and identically distributed as the test data then a trained model can make an accurate predictions for new samples of data. Conventional machine learning has a strong dependence on massive amounts of training data which are domain specific to understand their latent patterns. In contrast, Domain adaptation and Transfer learning methods are sub-fields within machine learning that are concerned with solving the inescapable problem of insufficient training data by relaxing the domain dependence hypothesis. In this contribution, this issue has been addressed and by making a novel combination of both the methods we develop a computationally efficient and practical algorithm to solve boundary value problems based on nonlinear partial differential equations. We adopt a meshfree analysis framework to integrate the prevailing geometric modelling techniques based on NURBS and present an enhanced deep collocation approach that also plays an important role in the accuracy of solutions. We start with a brief introduction on how these methods expand upon this framework. We observe an excellent agreement between these methods and have shown that how fine-tuning a pre-trained network to a specialized domain may lead to an outstanding performance compare to the existing ones. As proof of concept, we illustrate the performance of our proposed model on several benchmark problems.
ASJC Scopus Sachgebiete
- Informatik (insg.)
- Software
- Mathematik (insg.)
- Modellierung und Simulation
- Ingenieurwesen (insg.)
- Allgemeiner Maschinenbau
- Informatik (insg.)
- Angewandte Informatik
Zitieren
- Standard
- Harvard
- Apa
- Vancouver
- BibTex
- RIS
in: Engineering with computers, Jahrgang 38, Nr. 5, 10.2022, S. 4569-4588.
Publikation: Beitrag in Fachzeitschrift › Artikel › Forschung › Peer-Review
}
TY - JOUR
T1 - Domain adaptation based transfer learning approach for solving PDEs on complex geometries
AU - Chakraborty, Ayan
AU - Anitescu, Cosmin
AU - Zhuang, Xiaoying
AU - Rabczuk, Timon
N1 - Funding Information: AC acknowledge the support of the ERC Starting Grant no: 802205.
PY - 2022/10
Y1 - 2022/10
N2 - In machine learning, if the training data is independently and identically distributed as the test data then a trained model can make an accurate predictions for new samples of data. Conventional machine learning has a strong dependence on massive amounts of training data which are domain specific to understand their latent patterns. In contrast, Domain adaptation and Transfer learning methods are sub-fields within machine learning that are concerned with solving the inescapable problem of insufficient training data by relaxing the domain dependence hypothesis. In this contribution, this issue has been addressed and by making a novel combination of both the methods we develop a computationally efficient and practical algorithm to solve boundary value problems based on nonlinear partial differential equations. We adopt a meshfree analysis framework to integrate the prevailing geometric modelling techniques based on NURBS and present an enhanced deep collocation approach that also plays an important role in the accuracy of solutions. We start with a brief introduction on how these methods expand upon this framework. We observe an excellent agreement between these methods and have shown that how fine-tuning a pre-trained network to a specialized domain may lead to an outstanding performance compare to the existing ones. As proof of concept, we illustrate the performance of our proposed model on several benchmark problems.
AB - In machine learning, if the training data is independently and identically distributed as the test data then a trained model can make an accurate predictions for new samples of data. Conventional machine learning has a strong dependence on massive amounts of training data which are domain specific to understand their latent patterns. In contrast, Domain adaptation and Transfer learning methods are sub-fields within machine learning that are concerned with solving the inescapable problem of insufficient training data by relaxing the domain dependence hypothesis. In this contribution, this issue has been addressed and by making a novel combination of both the methods we develop a computationally efficient and practical algorithm to solve boundary value problems based on nonlinear partial differential equations. We adopt a meshfree analysis framework to integrate the prevailing geometric modelling techniques based on NURBS and present an enhanced deep collocation approach that also plays an important role in the accuracy of solutions. We start with a brief introduction on how these methods expand upon this framework. We observe an excellent agreement between these methods and have shown that how fine-tuning a pre-trained network to a specialized domain may lead to an outstanding performance compare to the existing ones. As proof of concept, we illustrate the performance of our proposed model on several benchmark problems.
KW - Domain adaptation
KW - Navier–Stokes equations
KW - NURBS geometry
KW - Transfer learning
UR - http://www.scopus.com/inward/record.url?scp=85130573733&partnerID=8YFLogxK
U2 - 10.1007/s00366-022-01661-2
DO - 10.1007/s00366-022-01661-2
M3 - Article
AN - SCOPUS:85130573733
VL - 38
SP - 4569
EP - 4588
JO - Engineering with computers
JF - Engineering with computers
SN - 0177-0667
IS - 5
ER -