[논문 리뷰] Car Drag Coefficient Prediction from 3D Point Clouds Using a Slice-Based Surrogate Model

U. Singh, Absaar Ali외 1명|arXiv (Cornell University)|2026. 01. 05.

Model Reduction and Neural Networks인용 수 0

한 줄 요약

본 논문은 PointNet2D 및 Bi-LSTM으로 인코딩된 순차적 2D 단면을 이용하여 3D 자동차 점군으로부터 드래그 계수(Cd)를 예측하는 경량 슬라이스 기반 대리 모델을 제안하며, 소비자용 GPU에서 높은 정확도와 빠른 추론을 달성한다.

ABSTRACT

The automotive industry's pursuit of enhanced fuel economy and performance necessitates efficient aerodynamic design. However, traditional evaluation methods such as computational fluid dynamics (CFD) and wind tunnel testing are resource intensive, hindering rapid iteration in the early design stages. Machine learning-based surrogate models offer a promising alternative, yet many existing approaches suffer from high computational complexity, limited interpretability, or insufficient accuracy for detailed geometric inputs. This paper introduces a novel lightweight surrogate model for the prediction of the aerodynamic drag coefficient (Cd) based on a sequential slice-wise processing of the geometry of the 3D vehicle. Inspired by medical imaging, 3D point clouds of vehicles are decomposed into an ordered sequence of 2D cross-sectional slices along the stream-wise axis. Each slice is encoded by a lightweight PointNet2D module, and the sequence of slice embeddings is processed by a bidirectional LSTM to capture longitudinal geometric evolution. The model, trained and evaluated on the DrivAerNet++ dataset, achieves a high coefficient of determination (R^2 > 0.9528) and a low mean absolute error (MAE approx 6.046 x 10^{-3}) in Cd prediction. With an inference time of approximately 0.025 seconds per sample on a consumer-grade GPU, our approach provides fast, accurate, and interpretable aerodynamic feedback, facilitating more agile and informed automotive design exploration.

연구 동기 및 목표

• Motivate fast, accurate Cd prediction in early automotive design to replace costly CFD/wind tunnel evaluations.
• Propose a 3D geometry representation as an ordered sequence of 2D cross-sectional slices along the streamwise axis.
• Develop a lightweight architecture combining PointNet2D per-slice encodings with a Bi-LSTM to model longitudinal geometry evolution.
• Demonstrate competitive accuracy with state-of-the-art surrogates on the DrivAerNet++ dataset while maintaining low computational cost and interpretability.

제안 방법

• Convert each 3D car point cloud into an ordered sequence of 80 2D cross-sectional slices along the X (streamwise) axis.
• Encode each 2D slice with a lightweight PointNet2D that outputs a 256-dim embedding per slice.
• Process the 80-slice embeddings with a 2-layer Bi-LSTM (256 hidden per direction) to obtain a 512-dim car-level representation.
• Regress Cd from the Bi-LSTM output using a 3-layer MLP (512 → 256 → 64 → 1) with ReLU activations and 0.3 dropout.
• Train with Smooth L1 loss (Huber, beta=1.0) using the Adam optimizer (lr=1e-4) for 100 epochs on an RTX 4060 laptop.
• Batch size is 4 due to memory constraints from the large per-slice point count (Mmax=6500).

Figure 2 : Visualization of 80 streamwise (X-axis) cross-sectional slices extracted from a vehicle’s point cloud. This grid displays the individual 2D point sets, sequentially arranged to highlight the progression of the vehicle’s contour from front to back, as used for feature extraction.

실험 결과

연구 질문

• RQ1 Can a slice-based sequential representation match or exceed the accuracy of existing 3D surrogates for Cd prediction?
• RQ2 How does the proposed model compare to state-of-the-art methods in terms of R^2 and MAE on DrivAerNet++?
• RQ3 Is the approach computationally efficient enough for real-time feedback on consumer-grade hardware?
• RQ4 Does the model offer interpretability by linking drag predictions to specific longitudinal slices?

주요 결과

모델	데이터셋 하위집합	MSE (1e-5)	MAE (1e-3)	MaxAE (1e-2)	R^2
PointNet2D+BiLSTM (Ours)	DrivAerNet++ (1200)	6.50	6.046	4.50	0.9528
TripNet [2]	DrivAerNet++ (1200)	9.10	7.17	7.70	0.957
RegDGCNN [4] a	DrivAerNet++ (1200)	14.20	9.31	12.79	0.641
PointNet [12] a	DrivAerNet++ (1200)	14.90	9.60	12.45	0.643

• Achieves R^2 = 0.9528 and MAE = 6.046e-3 on the DrivAerNet++ test set.
• Inferential speed of ~0.025 seconds per sample on an RTX 4060 Laptop GPU.
• Parameter count of ~2.79 million, indicating a lightweight model.
• Bi-LSTM effectively captures front-to-rear geometric evolution to predict Cd from sequence embeddings.
• Performance is competitive with TripNet while using simpler architecture and fewer parameters.
• Error analysis shows predictions are tightly clustered around true Cd with a unimodal error distribution.

Figure 3 : Complete model architecture of the proposed sequential slice-based drag prediction model. 3D point clouds are sliced. Each slice is encoded by PointNet2D. The sequence of embeddings is processed by a Bi-LSTM and an MLP regresses $C_{d}$ .