Flow-Conditioned Adapter (FCA)

• Flow-Conditioned Adapter (FCA) is a lightweight, residual affine-modulation module that adapts geometry-driven neural representations to varying aerodynamic flow conditions.
• It is integrated within the AdaField framework using a SAPT backbone, where only the FCA parameters are fine-tuned to maintain pre-trained knowledge and prevent overfitting.
• The design enables continuous, sample-efficient transfer across aerodynamic subdomains by conditioning features with task-specific flow variables to improve prediction accuracy.

A Flow-Conditioned Adapter (FCA) is a lightweight, residual affine-modulation module designed for efficient adaptation of geometry-driven neural representations to changing aerodynamic flow conditions in surface pressure field modeling tasks. FCA operates as an integral part of AdaField, a framework which leverages pre-training on large-scale public datasets and fine-tunes only a small set of parameters to enable rapid, sample-efficient transfer to new aerodynamic subdomains characterized by scarce computational fluid dynamics (CFD) data. The FCA module is implemented as a plug-in after each vector self-attention block in the Semantic Aggregation Point Transformer (SAPT) backbone, injecting flow-condition information into dense point cloud features and enabling conditional recalibration while keeping the majority of the model weights frozen (Zou et al., 12 Jan 2026).

1. Role and Purpose in Aerodynamic Model Adaptation

The FCA addresses the domain adaptation challenge posed by substantial differences in flow conditions—such as free-stream velocity, crosswind speed, Mach number, and angle of attack—that exist across modes of transportation (e.g., cars, trains, aircraft). In AdaField, the SAPT backbone efficiently extracts geometry-driven features from dense surface point clouds, but these geometric descriptors do not by themselves account for variations caused by flow regime changes. FCAs enable explicit conditioning of the model’s geometric representation with task-specific flow variables, functioning as learnable adapters that steer point-wise features in response to new flow regimes.

During pre-training (e.g., on DrivAerNet++), SAPT and all FCA layers are trained jointly to embed both geometric and flow-specific knowledge. For transfer to subdomains with limited data, the SAPT backbone—including all self-attention, feed-forward network (FFN), and aggregation parameters—remains frozen, and only the FCA modules are fine-tuned. This parameter-efficient strategy prevents catastrophic forgetting and overfitting to the small datasets typical of highly specialized aerodynamic applications, while allowing flexible, flow-conditioned recalibration.

2. FCA Module Architecture and Integration

The AdaField architecture alternates between vector self-attention blocks (Point Transformer style), FCA modules, and up/down-sampling via semantic aggregation or k-nearest neighbor interpolation in a U-Net paradigm. Each FCA block processes output features from the preceding point transformer and consists of several sub-blocks:

• Input Projection (Pin): Maps $D$-dimensional input features $x$ to a reduced $d$-dimensional adapter space via a Linear → LayerNorm → GELU sequence.
• Flow MLP: A two-layer MLP accepts the flow-condition vector $C \in \mathbb{R}^{D_f}$ and outputs channel-wise scale $\sigma \in \mathbb{R}^{d}$ and bias $\mu \in \mathbb{R}^{d}$.
• Modulation: Executes a channel-wise affine transformation, $u' = u \odot \sigma + \mu$, where $u = \text{Pin}(x)$ and $\odot$ denotes broadcasted multiplication.
• Output Projection (Pout): Returns to the original $D$-dimensional feature space via LayerNorm → GELU → Linear.
• Residual Addition: The transformed features are added back to the original input as $x \leftarrow x + y$.

This process is repeated after each transformer layer, allowing distinct flow-conditioned recalibration at multiple abstraction levels.

3. Formal Mathematical Description

Let $x \in \mathbb{R}^{N \times D}$ be the point-wise features after a Point Transformer block, and $C \in \mathbb{R}^{D_f}$ the flow condition vector. The operations of an FCA are as follows:

$$\begin{align*} (\sigma, \mu) &= \text{MLP}(C), \quad \sigma, \mu \in \mathbb{R}^{d} \ u &= \text{Pin}(x) \in \mathbb{R}^{N \times d} \ u' &= u \odot \sigma + \mu \ y &= \text{Pout}(u') \in \mathbb{R}^{N \times D} \ x_{\text{out}} &= x + y \end{align*}$$

where:

• $$\text{Pin}(x) = \text{GELU}(\text{LN}(x W_{\text{in}} + b_{\text{in}}))$$
• $$\text{Pout}(u') = \text{LN}(\text{GELU}(u' W_{\text{out}} + b_{\text{out}})) W'_{\text{out}} + b'_{\text{out}}$$
• $$\text{MLP}(C) = W_2 \cdot \text{GELU}(W_1 C + b_1) + b_2$$, producing $2d$ values that are split into $\sigma$ and $\mu$.

4. Parameterization and Fine-Tuning Regime

Quantitative parameter scaling in the default configuration (with $D=128$, $d=32$, $D_f=2$) is as follows:

FCA Component	Parameters (per block)	Notes
Pin	$(128 \times 32) + 32$ + LN params ≈4.2K	Linear + LayerNorm + GELU
Pout	$(32 \times 128) + 128$ + LN params ≈4.3K	LayerNorm + GELU + Linear
Flow MLP	$(2 \times 64 + 64) + (64 \times 64 + 64)+2d ≈9.8K$	2-layer MLP + split
Total/FCA	≈18.3K
12 FCAs total	≈220K	With 12 SAPT layers
SAPT Backbone	~50M	All frozen during adaptation

During domain adaptation, only the $\approx220$K adapter parameters (∼0.5% of the full network) are updated, with all other ∼50M SAPT parameters frozen. This enables parameter-efficient adaptation and guards against overfitting.

5. Training and Adaptation Workflow

The learning and adaptation sequence employs the following methodology:

• Pre-Training: The entire SAPT + FCA network is trained end-to-end on near-exhaustive, large-scale datasets (e.g., DrivAerNet++), with physics-informed data augmentation (PIDA) to increase coverage of object scale and velocity variation. Optimization uses Adam (β₁=0.9, β₂=0.999), learning rate $1 \times 10^{-4}$, batch size 2 (≈32K points), and 200 epochs. The loss function is Mean Squared Error (MSE) between predicted and ground-truth non-dimensional surface pressure coefficient $C_p$.
• Adaptation: For subdomains with scarce CFD data (e.g., high-speed rail or specific aircraft geometries), SAPT and PIDA modules are frozen, and only FCA parameters are fine-tuned. Learning rate is reduced to $5 \times 10^{-5}$; convergence is typically achieved within ≈20 epochs, even with as few as 1–30 training samples.

6. Empirical Impact and Sample Efficiency

Fine-tuning only the FCAs delivers substantial improvements in sample efficiency and generalization:

• On the train dataset for pressure field modeling, fine-tuning FCAs lowers MSE from $1.79 \times 10^{-2}$ (training full model from scratch) to $0.99 \times 10^{-2}$.
• For an aircraft wing, MSE drops from $16.63 \times 10^{-2}$ to $11.46 \times 10^{-2}$ under the same protocol.
• A single fine-tuning sample in the train domain yields MSE below $1.2 \times 10^{-2}$ (vs. $>1.7 \times 10^{-2}$ from-scratch).
• With 20% of the scarce wing data (28/140 samples), MSE with FCA adaptation is within 5% of the full-data model.

The sample-efficiency curve further demonstrates that FCA fine-tuning rapidly approaches full-model accuracy as a function of transfer set size (Zou et al., 12 Jan 2026).

7. Mechanisms Underlying Generalization Gains

The improvements delivered by FCA design stem from several key factors:

• Geometry–Flow Decoupling: The SAPT backbone encodes comprehensive, flow-agnostic geometric information, while FCA modules perform conditional, data-driven modulation, steering features adaptively in the flow condition space.
• Parameter Efficiency: By restricting adaptation to a small set of parameters (>99% frozen), AdaField mitigates overfitting risk and preserves prior aerodynamic knowledge from the pre-training phase.
• Continuous Conditioning: The adapter’s learned $\sigma, \mu$ can interpolate continuously across flow conditions (e.g., $v=30$ m/s to $v=100$ m/s, or $\mathrm{Ma}=0.2\rightarrow0.8$), supporting smooth generalization and preventing extrapolation failures.
• Residual Correction: Each FCA block employs a residual design—adding only a small learned correction $y$ to the pre-trained feature $x$—ensuring that in out-of-distribution conditions the model defaults gracefully to its initial (pre-trained) prediction rather than producing erroneous outputs.

Collectively, these attributes enable FCAs to achieve a practical compromise between rigid weight freezing (inflexible to new flow regimes) and full end-to-end retraining (computationally expensive and data-hungry), supporting robust generalization and efficient transfer in data-scarce aerodynamic modeling settings (Zou et al., 12 Jan 2026).