High-strength metal alloys achieve their performance via careful control of precipitates and solutes. The nucleation, growth, and kinetics of precipitation, and the resulting mechanical properties, are inherently atomic scale phenomena, particularly during early-stage nucleation and growth. Atomistic modeling using interatomic potentials is a desirable tool for understanding the detailed phenomena involved in precipitation and strengthening, which requires length and time scales far larger than those accessible by first-principles methods. Current interatomic potentials for alloys are not, however, sufficiently accurate for such studies. Here, a family of neural-network potentials (NNPs) for the Al-Cu system are presented as a first example of a machine-learning potential that can achieve near-first-principles accuracy for many different metallurgically-important aspects of this alloy. High fidelity predictions of intermetallic compounds, elastic constants, dilute solid-solution energetics, precipitate/matrix interfaces, generalized stacking fault energies and surfaces for slip in matrix and precipitates, antisite defect energies, and other quantities, are shown. The NNPs also captures the subtle entropically-induced transition between $\theta \prime$ and $\theta$ at temperatures around 600K. Many comparisons are made with the state-of-the-art Angular-Dependent Potential for Al-Cu, demonstrating the significant quantitative benefit of a machine-learning approach. A preliminary kinetic Monte Carlo study shows the NNP to predict the emergence of GP zones in Al-4at%Cu at T=300K in agreement with experiments. These studies show that the NNP has significant transferability to defects and properties outside the structures used to train the NNP, but also shows some errors highlighting that the use of any interatomic potential requires careful validation in application to specific metallurgical problems of interest.

中文翻译：

---

冶金学的机器学习I.铝铜的神经网络潜力

高强度金属合金通过仔细控制析出物和溶质实现其性能。析出的成核，生长和动力学以及由此产生的机械性能，本质上是原子尺度的现象，尤其是在早期成核和生长过程中。使用原子间势能进行原子建模是了解与降水和强化有关的详细现象的理想工具，这需要长度和时间尺度远大于第一原理方法可达到的尺度。然而，合金的当前原子间电势对于此类研究而言还不够准确。这里，提出了一个用于Al-Cu系统的神经网络电位族（NNP），作为该合金许多冶金学重要方面可以实现接近第一原理精度的机器学习电位的第一个示例。显示了金属间化合物的高保真度预测，弹性常数，稀固溶体能学，沉淀物/基体界面，广义堆垛层错能以及基体和沉淀物中的滑动面，反位缺陷能及其他数量。NNPs还捕获了由熵引起的微妙的过渡 给出了广义的堆垛层错能以及基体和沉淀中的滑动面，反位缺陷能及其他数量。NNPs还捕获了由熵引起的微妙的过渡 给出了广义的堆垛层错能以及基体和沉淀中的滑动面，反位缺陷能及其他数量。NNPs还捕获了由熵引起的微妙的过渡$\theta \prime$ 和 $\theta$在约600K的温度下。使用最新的Al-Cu角相关电位进行了许多比较，证明了机器学习方法的显着定量收益。初步的动力学蒙特卡洛研究表明NNP可以预测在T = 300K时Al-4at％Cu中GP区域的出现，与实验一致。这些研究表明，NNP在用于训练NNP的结构之外具有向缺陷和特性的显着转移能力，但也显示出一些错误，突出表明使用任何原子间电势都需要在验证特定的冶金问题时进行仔细验证。