Q1：**文章提出的工程问题是什么？

    有什么实际工程价值？**

Present a new trial to reproduce soil stress-strain behaviour

First discovered a new pehnomeon “bias at low stress levels” for nn methods

Q2：文章提出的学术问题是什么？

     有什么新的学术贡献？

The capacity is outperforms the feed forward and feedback neural networks

To reduce the soil’s stress-strain behaviour

Q3：文章提出的技术路线是什么？

     有什么改进创新之处？

three steps:

data preparation

architecture determination

optimistisation

Q4：文章是如何验证和解决问题的？

1. Introduction stress-strain of behaviours soils show high nonlinearity soli behaviour model base on mechanical hypotheses(elasticity elastoplasticity hypoplasticity) one model only for a special soil and difficult to determine NN - FFNNs -FBNNs - NANNs -RNNs LSTM can learn both the long and short_term influences
2. Methofology

• 2.1 LSTM deep-learning network
• 2.2.1 data preparation: component of input data , data be normalised 0-1
• 2.2.2 architecture determination: adptive determination method , empirical method
• 2.2.3 optimisation method: evolutionary algorithms gradient decent(low computer cost) :batch gradient (small datasets)/ stochastic gradient descent algorithms

1. Experiments and results

numerical to verifying the improvement of LSTM in constrast to nn

laboratory dataset to comfirm the capacity os LSTM on real soil behaviours

• 3.1.1 numerical experiment Plaxis software; drained traixial ;100 training dataset 27testing dataset ;each 30-60steps up to 3000-6000 pairs
• 3.1.2 laboratory test sourced from Lee and Seed
• 3.2 modelling details three model : FFNN,FBNN,LSTM
• 3.3.1 numerical experiments LSTM outperform the feedback and feedforward models in convergence rate and precision. feedback model outperformed the feedforward model LSTM was more effective / overwhelm than the other two

1. discussion

• 4.1 bias at low stress levels a high stess level corresponds to a higher soil strength the problem of bias at low stress levels should be examined in detail in future studies
• 4.2 complexity of models 229 weights of ff, 241 of fb, 985 of LSTM LSTM is worth the extra computer costs

1. conclusions

• a new approach model the stress-strain behaviours by LSTM (Octave soft ware)
• LSTM outperformed in the precision and convergence rate
• LSTM had excellent performance on the measured stress-strain bahaviours
• bias at low stress is a common problem

Q5：文章有什么可取和不足之处？

计算机方面的Q1期刊

Q6：文章对自身的研究有什么启发？

启发很多：

1. LSTM的结构组成
2. 误差的箱型图 boxplots
3. 雷达图的应用
4. numerical 的来源，Plaxis进行大量的模拟