基於物理引導神經運算子之三維空氣超音波預測與數位孿生建構

Abstract

空氣超音波陣列廣泛應用於非接觸式懸浮、物體辨識、指向性音響及觸覺合成等領域，然而這些應用的開發過程往往仰賴高成本且耗時的模擬與實驗，此外若欲結合人工智慧技術以提升系統效能，則需大量高品質資料作為訓練基礎，但資料的取得同樣面臨高昂代價與實作困難。為解決此一瓶頸，本研究提出一套基於物理引導神經算子（Physics-Informed Neural Operators, PINO）的數位孿生建構方法，針對空氣超音波陣列於三維非均質環境中的傳播行為進行模擬。我們於虛擬環境中建立傳感器陣列的數位對應模型，並訓練 PINO 模型以預測超音波在不同介質與空間分布下的三維時序聲場變化，該模型的結果能同時轉換成穩態聲場與感測器所接收之訊號，且其預測結果與真實物理行為高度一致。相較於傳統數值模擬方法，本研究所提出之 PINO 模型具備超過 20 倍的加速能力，並且能在相同硬體資源下模擬更大規模的三維場域，藉由此數位孿生平台，可快速進行多樣化的模擬實驗，亦可有效生成大量高品質資料，以支援 AI 模型之訓練與應用開發，進而大幅降低開發門檻與成本。

Airborne ultrasound arrays are widely utilized in applications such as non-contact levitation, object recognition, directional audio, and haptic feedback. However, the development of these applications often relies on high-cost and time-consuming simulations and experiments. Furthermore, integrating artificial intelligence to enhance system performance requires large volumes of high-quality training data, the acquisition of which is both expensive and labor-intensive. To address these challenges, this study proposes a digital twin construction framework based on Physics-Informed Neural Operators (PINO) to simulate the propagation behavior of airborne ultrasound arrays in three-dimensional heterogeneous environments. A digital counterpart of the sensor array is built within a virtual environment, and a PINO model is trained to predict the spatiotemporal evolution of the acoustic field across various media and spatial distributions. The model outputs both the steady-state sound field and the sensor-received signals, achieving high fidelity with realworld physical behavior. Compared to conventional numerical simulation methods, the proposed PINO model offers more than a 20-fold improvement in computational speed and enables the simulation of larger-scale 3D domains under the same hardware constraints. This digital twin platform facilitates rapid and diverse simulation experiments and enables the efficient generation of high-quality data to support AI model training and application development, thereby significantly reducing the development cost and complexity.