正文：
近日，丹麦哥本哈根大学的研究人员开发了一种利用深度学习解决晶体学相位问题的创新方法，这一成果发表在《Science》杂志上。该研究团队开发了名为PhAI的深度学习神经网络，通过数百万个人工晶体结构及其衍射数据的训练，能够生成准确的电子密度图。
晶体学是研究晶体物质结构的关键技术，通过X射线衍射等技术可以获得晶体的三维结构。然而，晶体学中的一个长期难题——相位问题，一直是困扰科学家们的一大挑战。传统的晶体学实验只能获得结构因子的振幅，而相位信息丢失，这限制了晶体学在更高分辨率下的应用。
PhAI方法通过深度学习技术，能够以仅2埃的分辨率解决相位问题，该分辨率相当于原子分辨率可用数据的10%到20%。这一突破性进展意味着，即使是在传统方法需要原子分辨率数据的条件下，PhAI方法也能够提供解决方案。
研究人员采用了一种数据驱动的方法，利用数百万个人造晶体结构及其衍射数据来训练神经网络。这些结构包括有机分子、金属有机晶体和无机晶体，涵盖了广泛的化学和材料科学领域。
PhAI神经网络的设计与训练，构建了一种名为PhAI的神经网络，它能够接受结构因子振幅并输出相应的相位值。该网络通过一系列的卷积和多层感知器块处理输入数据，最终输出预测的相位值。
经过测试，PhAI方法在计算真实晶体结构的衍射数据时表现出高准确率。研究人员共获得了2387个测试用例，证明了PhAI方法在解决真实结构问题上的有效性和实用性。
这一成果不仅为晶体学领域提供了新的解决方案，也为其他需要解决复杂相位问题的领域提供了新的研究思路。随着深度学习技术的不断进步，未来有望在更多科学领域中发挥作用。
英语如下：
News Title: “Deep Learning Breakthrough: Crystal Phase Problem Solved with Ease”
Keywords: Crystallography, Deep Learning, Science Journal
News Content:
Title: University of Copenhagen Develops Deep Learning Method to Solve Crystal Phase Problem
Recently, researchers at the University of Copenhagen have developed an innovative method to solve the crystal phase problem using deep learning, which has been published in the Science journal. The research team has developed a deep learning neural network named PhAI, which, after being trained on millions of artificial crystal structures and their diffraction data, can generate accurate electron density maps.
Crystallography is a key technique for studying the structures of crystalline materials, allowing the three-dimensional structure of crystals to be obtained through techniques such as X-ray diffraction. However, a longstanding challenge in crystallography – the phase problem – has been a significant hurdle for scientists. Traditional crystallographic experiments can only provide amplitude information of the structure factors, with phase information lost, which limits the application of crystallography at higher resolutions.
The PhAI method utilizes deep learning technology to solve the phase problem at a resolution of only 2 angstroms, which is approximately 10% to 20% of the atomic resolution data available. This breakthrough indicates that even when traditional methods require atomic resolution data, the PhAI method can provide solutions.
The researchers employed a data-driven approach, training the neural network with millions of artificial crystal structures and their diffraction data. These structures included organic molecules, metal-organic crystals, and inorganic crystals, covering a wide range of fields in chemistry and materials science.
The design and training of the PhAI neural network have constructed a neural network named PhAI, which can accept amplitude values of the structure factors and output the corresponding phase values. The network processes input data through a series of convolutional and multi-layer perceptron blocks, ultimately outputting predicted phase values.
Tests have shown that the PhAI method exhibits high accuracy when computing diffraction data for real crystal structures. The researchers have obtained 2,387 test cases, proving the effectiveness and practicality of the PhAI method in solving real structure problems.
This achievement not only provides a new solution for the crystallography field but also offers new research ideas for other areas that need to solve complex phase problems. With the continuous advancement of deep learning technology, it is expected to play a role in more scientific fields in the future.
【来源】https://www.jiqizhixin.com/articles/2024-08-08-3
Views: 1