-</pre></div></div><a class="reference internal image-reference"href=../../_images/5e17956440bff7601109d057f8fe381e746ab551a61d48a23690f9eae5edd4ad.png><img alt=../../_images/5e17956440bff7601109d057f8fe381e746ab551a61d48a23690f9eae5edd4ad.png src=../../_images/5e17956440bff7601109d057f8fe381e746ab551a61d48a23690f9eae5edd4ad.png style=width:400px> </a></div></div></section><section id=training-environment-correction><h4><strong>2.4 Training Environment Correction</strong><a title="Link to this heading"class=headerlink href=#training-environment-correction>#</a></h4><p>The <strong>DeePTB-SK</strong> module provides powerful environment-dependent modeling with symmetry-preserving neural networks. Based on the previously constructed <code class="docutils literal notranslate"><span class=pre>nnsk</span></code> model, we can further enhance the TB model’s descriptive ability by adding an environment-dependent component to overcome the accuracy limitations imposed by the two-center approximation. The model that incorporates environment dependence into the <code class="docutils literal notranslate"><span class=pre>nnsk</span></code> model is referred to as the <code class="docutils literal notranslate"><span class=pre>mix</span></code> model, and its expression is as follows: $$</p><div class="amsmath math notranslate nohighlight"id=equation-c6c34ac4-93e1-4469-bc82-8707be8e5396><span class=eqno>(1)<a title="Permalink to this equation"class=headerlink href=#equation-c6c34ac4-93e1-4469-bc82-8707be8e5396>#</a></span>\[\begin{equation} h^{\text{env}}_{ll^\prime{\zeta}} = h_{ll^\prime{\zeta}}(r_{ij}) \times \left[1+\Phi_{ll^\prime\zeta}^{o_i,o_j}\left(r_{ij},\mathcal{D}^{ij}\right)\right] \end{equation}\]</div><p>$<span class="math notranslate nohighlight">\( where \)</span>\mathcal{D}^{ij}<span class="math notranslate nohighlight">\( is the environment descriptor defined by the `embedding` keyword, and \)</span>\Phi_{ll^\prime\zeta}^{o_i,o_j}$ is the neural network that provides the environment correction prediction value.</p><p>To define the <code class="docutils literal notranslate"><span class=pre>mix</span></code> correction model, you need to provide the following keywords in the <code class="docutils literal notranslate"><span class=pre>model_options</span></code> section of the training input file:</p><ul class=simple><li><p><code class="docutils literal notranslate"><span class=pre>embedding</span></code>: The <code class="docutils literal notranslate"><span class=pre>method</span></code> here specifies the form of the atomic environment used in the <code class="docutils literal notranslate"><span class=pre>dptb</span></code> model. In this example, we use the <code class="docutils literal notranslate"><span class=pre>se2</span></code> form of descriptor similar to that used in <strong>DeePMD</strong>.</p></li><li><p><code class="docutils literal notranslate"><span class=pre>prediction</span></code>: The <code class="docutils literal notranslate"><span class=pre>method</span></code> specifies the prediction method of the model, which is set to <code class="docutils literal notranslate"><span class=pre>sktb</span></code> here. The <code class="docutils literal notranslate"><span class=pre>neurons</span></code> keyword specifies the size of the prediction network.</p></li><li><p><code class="docutils literal notranslate"><span class=pre>nnsk</span></code>: This section is consistent with the content in the <code class="docutils literal notranslate"><span class=pre>nnsk</span></code> model. The <code class="docutils literal notranslate"><span class=pre>freeze</span></code> option should be set to <code class="docutils literal notranslate"><span class=pre>true</span></code>, indicating that the trained SK parameters of the <code class="docutils literal notranslate"><span class=pre>nnsk</span></code> model are fixed, and only the neural network parameters of the environment-dependent part are trained. This fixing is crucial; otherwise, the initialization of the <code class="docutils literal notranslate"><span class=pre>mix</span></code> model may completely destroy the parameters of the <code class="docutils literal notranslate"><span class=pre>nnsk</span></code> model, leading to non-convergence during training.</p></li></ul><p>For example:</p><div class="highlight-json notranslate"><div class=highlight><pre><span></span><span class=w> </span><span class=nt>"model_options"</span><span class=p>:</span><span class=w> </span><span class=p>{</span>
0 commit comments