While Surface Heat Flow (HF) is an important constraint unveiling the Earth interior's thermal structure, estimates over the Iranian plateau are sparse. In the presence of sparse estimates, machine learning provides a statistical-based prediction of HF based on a supervised predictor trained in the far-field regions. Here, we imply the machine learning technique of Gradient Boosting Regression Tree (GBRT) which has been proved to be efficient for predicting HF projecting complexities and nonlinearities of input features into predicted HF. Our results provide a robust map of HF with resolution of one degree and uncertainty of up to ±6 mW/m over Iran and surrounding regions. The predicted HF has an average value and minimum standard deviation of 59 and 10 mW/m, respectively. The quality of the algorithm performance is 16%, indicated by normalized Root-Mean-Square Error (RMSE), and linear correlation of predicted HF with validation set is 97%. Total number of trees considerably prevents overfitting which is believed to be solely controllable by shrinkage factor, maximum tree depth and cross-validation scheme. The three most important features, having the highest influence on the output HF, are thermal Lithosphere-Asthenosphere Boundary (LAB), distance to volcanoes and distance to trenches. The extreme importance of LAB in HF prediction of Iran indicates that the lithospheric thermal structure is significantly controlled by lithospheric thickness in the Iranian plateau. Selection of petrologically and seismologically consistent LAB guarantees the precision of the predicted HF. Our results imply that high HF in central Iran is in agreement with extensive magmatism since the Paleozoic. Additionally, the high HF in Zagros keel (originally Proterozoic as the Zagros keel appears to be the Arabian plate front) indicates the tectonically active system of the Arabia-Eurasia collision zone, high likely, in the form of lithospheric mantle deformation.

中文翻译：

---

基于机器学习方法预测伊朗高原及周边地区的热流


虽然表面热流（HF）是揭示地球内部热结构的一个重要约束因素，但对伊朗高原的估计却很少。在存在稀疏估计的情况下，机器学习根据在远场区域训练的监督预测器提供基于统计的 HF 预测。在这里，我们指的是梯度提升回归树（GBRT）的机器学习技术，该技术已被证明对于预测 HF 将输入特征的复杂性和非线性投影到预测的 HF 中是有效的。我们的结果提供了伊朗及周边地区的可靠 HF 地图，其分辨率为 1 度，不确定性高达 ±6 mW/m。预测的 HF 平均值和最小标准偏差分别为 59 和 10 mW/m。以归一化均方根误差（RMSE）表示，算法性能质量为 16%，预测 HF 与验证集的线性相关性为 97%。树的总数大大防止了过度拟合，这被认为只能通过收缩因子、最大树深度和交叉验证方案来控制。对输出高频影响最大的三个最重要的特征是热岩石圈-软流圈边界 (LAB)、到火山的距离和到海沟的距离。 LAB在伊朗HF预测中的极端重要性表明，伊朗高原岩石圈热结构明显受岩石圈厚度控制。选择岩石学和地震学一致的 LAB 保证了 HF 预测的精度。我们的结果表明伊朗中部的高高频与古生代以来广泛的岩浆作用一致。 此外，扎格罗斯龙骨中的高高频（最初是元古代，因为扎格罗斯龙骨似乎是阿拉伯板块前缘）表明阿拉伯-欧亚大陆碰撞带的构造活跃系统，很可能以岩石圈地幔变形的形式存在。