BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Physics-informed Neural Networks for Simultaneous Surrogate Modell
 ing and Aerodynamic Optimization  - Dr. Yubiao Sun
DTSTART:20220513T113000Z
DTEND:20220513T123000Z
UID:TALK172946@talks.cam.ac.uk
CONTACT:99736
DESCRIPTION:Optimizing multiple variable problems is an exceedingly diffic
 ult task due to the curse of dimensionality. This is particularly true for
  airfoil shape optimization as remeshing or deformation of existing meshes
  is required\, which is computationally expensive. Our study aims to overc
 ome this challenge by introducing a deep learning-based framework that can
  handle multiple optimization problems efficiently. The essence of this fr
 amework is using surrogate modelling to produce high-fidelity solutions an
 d then perform gradient-based optimizations for high-dimensional problems.
  The starting point is to use PINN to construct surrogate models that outp
 ut flow fields for airfoils of varied configurations. The key feature of P
 INN is the incorporation of physical problem description\, including the g
 overning laws of physics\, domain geometry\, and boundary conditions\, whi
 ch enables neural networks to solve underlying differential equations (e.g
 .\, Navier--Stokes equations) as a learning problem. Thus\, the surrogate 
 models can efficiently generate flow fields as no labelled training data f
 rom a separate high-fidelity simulation is required. More importantly\, we
  extend the employed surrogate models by including design parameters as in
 puts to PINN. In the optimization process\, a quasi-Newton algorithm is us
 ed and further accelerated by automatic differentiation\, a popular algori
 thm designed to efficiently compute the gradients of objective functions w
 ith respect to design variables. Two examples have been presented to demon
 strate the feasibility of using PINN-based surrogate modelling for aerodyn
 amic optimization\, both single parameter and multiple parameter problems.
  The proposed method is straightforward to implement and computationally e
 fficient\, providing a promising alternative for computationally intensive
  optimization problems.  
LOCATION:CUED\, LR3B
END:VEVENT
END:VCALENDAR