Vision transformer-based physics informed CFD prediction of axial fans with self-supervised contrastive learning for enhanced geometric sensitivity

Abstract

In recent years, deep learning has emerged as a transformative approach in aerodynamic modeling, offering efficient alternatives to Computational Fluid Dynamics methods, particularly for design-intensive applications in turbomachinery. This thesis proposes a physics-informed deep learning framework for predicting aerodynamic performance parameters across multi-stage fan geometries, leveraging the Vision Transformer (ViT) architecture for its capacity to model complex spatial patterns and geometric variability. The proposed architecture is composed of an encoder-decoder structure specifically designed for turbomachinery geometries. The encoder utilizes a ViT backbone, where input geometries are divided into non-overlapping patches and embedded into high-dimensional feature vectors via convolutional patch embedding. Positional encoding is added to preserve spatial structure, followed by multiple self-attention layers that extract long-range geometric dependencies. To account for varying stage configurations, attention masking and a learnable inlet station token are integrated into the encoder pipeline. The decoder begins with an attention-based upsampler that increases the spatial resolution of the latent representation, followed by a boundary conditioning module that modulates the geometric features using Feature-wise Linear Modulation (FiLM) for inlet conditions and cross-attention for RPM and outlet pressure. The final stage of the decoder, referred to as the projector, consists of multi-layer perceptrons that reduce the feature dimensionality and map the encoded information to the target output space. The architecture is carefully designed to ensure that both spatial and physical structures are retained throughout the transformation. To enhance geometric sensitivity and avoid overfitting to global input–output mappings, the encoder is first pretrained using contrastive learning with the NT-Xent loss. This step enables the model to distinguish subtle geometric variations, which are critical in turbomachinery design. The decoder is then fine-tuned using supervised learning, guided by a combination of a percentage-based loss and two physics-based losses: a continuity loss enforcing mass conservation, and an isentropic loss ensuring thermodynamic consistency between predicted flow variables. The model is evaluated across a comprehensive CFD dataset covering multiple speedlines and six distinct fan designs, including variations in metal angle and hub radius. The results demonstrate that the model successfully captures both radial and streamwise flow behaviors, performs reliably across off-design operating conditions, and accurately simulates performance trends in response to design modifications. Notably, even with subtle geometric changes, the model reflects physically consistent shifts in aerodynamic quantities. This work highlights the efficacy of integrating self-supervised learning with physics-based constraints in achieving generalizable, computationally efficient aerodynamic models. The approach offers significant promise as a surrogate model for accelerating turbomachinery design cycles and supporting early-stage performance evaluations under varying geometric and operational scenarios.