• ISSN 2097-1893
  • CN 10-1855/P

基于深度神经网络的南极洲大地热流预测

Prediction of geothermal heat flow in Antarctica based on deep neural network

  • 摘要: 南极冰盖对全球气候系统的调控作用愈发显著,其中大地热流(GHF)作为基底热状态的关键指标,了解其分布特征对研究动力学演化具有重要意义. 南极洲GHF测量受限于极寒地质条件与高昂钻孔成本,已知的热流点仅有52个. 基于地质地球物理数据估计的热流分布存在较大不确定性,无法解释基底的非线性传热过程. 深度学习算法可以捕捉热流与不同数据特征的耦合关系,对稀疏数据插值重构具有较高的预测精度. 本研究构建了一种深度神经网络(DNN)与高维不确定性量化框架,用于预测南极洲GHF分布. 通过将4000多个全球热流观测点及22类地质地球物理特征参数(地壳厚度、重磁异常、莫霍面等)作为训练数据,对澳大利亚与南极已知热流点进行了测试对比. 不确定性量化方法通过相关性、灵敏度和主成分分析重组训练集,结合DNN模型评估南极洲最佳热流分布. 预测结果显示南极洲热流范围为24~103 \mathrmm\mathrmW/\mathrmm^2 ,平均值为60.1 \mathrmm\mathrmW/\mathrmm^2 ,与已知热流误差为10.05%,其中甘布尔采夫冰下山脉表现出显著高热流异常(50~74 \mathrmm\mathrmW/\mathrmm^2 ). 通过不确定量化结果证实甘布尔采夫冰下山脉的高预测热流具有较高置信度. 该模型解译了南极地壳热结构的空间异质性特征,预测结果可用于南极尺度的冰盖动力学或热力学模拟,为冰盖运动、冰下湖分布等研究提供了可靠的数据支持.

     

    Abstract: The Antarctic ice sheet plays an increasingly significant regulating role in the global climate system, where geothermal heat flow (GHF) serves as a critical indicator of basal thermal conditions. Understanding its spatial distribution is essential for studying its dynamic evolution. Current GHF measurements in Antarctica remain limited to merely 52 data points due to extreme cold geological conditions and prohibitive drilling costs. Existing GHF estimates based on geological and geophysical data exhibit substantial uncertainties and fail to account for non-linear heat transfer processes in the bedrock. Deep learning algorithms demonstrate superior predictive capabilities for sparse data interpolation by capturing coupled relationships between GHF and multi-source data features. We develop a deep neural network (DNN) integrated with a high-dimensional uncertainty quantification framework to predict Antarctic GHF distribution. Utilizing over 4000 global GHF observations and 22 geological or geophysical features (including crustal thickness, gravity/magnetic anomalies, and Moho depth) as training data, we validate the model against known GHF points in Australia and Antarctica. The uncertainty quantification method reorganizes the training dataset through correlation, sensitivity, and principal component analysis, combined with DNN model to determine optimal GHF distribution. Prediction results reveal the GHF range of 24-103 \mathrmm\mathrmW/\mathrmm^2 with a mean value of 60.1 \mathrmm\mathrmW/\mathrmm^2 , showing a 10.05% error relative to measured data. The Gamburtsev Subglacial Mountains exhibit significant high GHF anomalies (50-74 \mathrmm\mathrmW/\mathrmm^2 ), Uncertainty quantification confirms high confidence levels for these GHF predictions. Our model successfully interprets spatial heterogeneity characteristics of Antarctic crustal thermal structure, and the prediction results can be further applied to model Antarctic-scale ice sheet dynamics or thermodynamics, provide reliable data support for studies on ice sheet motion and subglacial lake distribution.

     

/

返回文章
返回