Research on a three-dimensional radiation field reconstruction algorithm based on an improved 3D CNN.

Area of Science:

• Computational physics
• Artificial intelligence
• Radiation detection

Background:

• Accurate reconstruction of radiation fields is crucial for various applications.
• Sparse data presents significant challenges in achieving high-fidelity field reconstruction.
• Existing methods often struggle with efficiency and precision when data is limited.

Purpose of the Study:

• To develop and validate an improved three-dimensional convolutional neural network (3D CNN) for reconstructing radiation fields from sparse data.
• To assess the network's performance across different radiation environments and data sampling densities.
• To demonstrate the efficacy of a self-attention integrated CNN for interpolating and generating complete radiation distribution grids.

Main Methods:

• Consolidation of sparse radiation data points into structured three-dimensional matrices.
• Application of a self-attention integrated 3D CNN for data interpolation and grid generation.
• Experimental validation using randomly sourced radiation in unshielded and shielded environments, and with refined grid configurations.

Main Results:

• In unshielded environments, 5% data sampling achieved a 4% average relative error.
• In shielded settings, 7% data sampling resulted in approximately 11% error.
• A 2% sampling rate in refined grids limited the error to 6.58%.

Conclusions:

• The improved 3D CNN demonstrates high effectiveness for precise three-dimensional radiation field reconstruction.
• The model successfully interpolates radiation distribution grids even with highly sparse data.
• This approach offers a robust solution for accurate radiation mapping in data-scarce scenarios.