BP人工神经网络用于除草剂活性预测

The Prediction for the Activity of Herbicide Based on Back-Propagation Artificial Neural Network

采用AM1方法研究了10种新型的磺酰脲类除草剂的电子结构,并以原子的Mulliken净电荷和除草剂在不同浓度(100,10mg/L)下对油菜、稗草两种作物的根、茎部位的抑制率为训练样本集,构造并训练得到具有活性预测能力的BP神经网络.结果表明,该BP网络不仅能对训练样本很好拟合,亦能对未知化合物的活性作出很好的预测.

The geometry of 10 herbicides was optimized by using AM1 method. An BP neural network (30-15-4) was developed to predict the activity of herbicide. The Mulliken net charges of 29 atoms and the concentration (100,10 mg/L) of herbicides were selected as the 30 input values of network, and the activity of herbicides as the 4 output values of network. Simulation and prediction results indicate that the choice of the characters is reasonable, and the method based on BP is effective.