Machine Learning for Designing Mixed Metal Halides for Efficient Ammonia Separation and Storage

Abstract

Metal halide salts have been proposed as promising absorbents for high-temperature separation of ammonia in the intensified Haber−Bosch process. A challenge for the widespread application of metal halides in such applications is the regeneration of absorbents, which requires energy-intensive temperature and pressure swings. New mixed metal halides could be prepared from the pure metal halides, offering a huge number of combinations that could be used to optimize the operating costs. To achieve this goal, knowledge of the deammoniation reaction energies is needed to direct the swing cycles efficiently. Computational methods offer excellent opportunities to obtain these data in an economic way. This work presents a combined density functional theory (DFT) and machine learning (ML) approach to manage these large amounts of mixing possibilities by efficiently and accurately predicting the required deammoniation reaction energies. Mixed metal halides (MMHs) of Mg, Ca, Cl, and Br [Mg x Ca 1−x Cl y Br 2−y (NH 3 ) n , x = 0−1, y = 0− 2, and n = 6, 2, 1 and 0] were selected with deammoniation steps n following the coordination numbers 6 → 2→ 1 → 0. To construct a data-efficient and computationally inexpensive ML approach, we (i) developed an efficient interpolation scheme based on DFT calculations for the structures of the MMHs based on those of the pure compounds, (ii) performed fixed-structure DFT calculations for creating training and test sets, and (iii) analyzed the importance of the features to be used in the ML procedure. Remarkably, our ML approach required a very small training set of 45 cases from a total of 4096 cases per deammoniation reaction step to achieve satisfactory predictions with chemical accuracy. Deammoniation energies were calculated with a standard deviation of better than ±0.1 kcal/mol for the step 6 → 2 and up to ±0.8 kcal/mol for the step 1 → 0. This approach has a potential to be applicable to a broad range of metal halides composed of metals from alkaline earth elements and/or 3d and 4d metals.

AI in computational chemistry through the lens of a decade-long journey

“…Remarkably, we were able to train ML achieving chemical accuracy for predicting the deammoniation energies which is the key property of such materials. 38 The crucial finding was that we can get much better models when we train the models for intermediate QM properties such as ionic polarizabilities and charges which are then fed into the ML model predicting the final target property. 38 Similarly, we used ML to learn the two-photon absorption on relatively small experimental data (less than thousand data points) and were able to predict absorption strengths for new compounds which were verified in the lab; this approach has the potential to be used for screening better materials for, e.g., upconverted laser or 3D printing.…”

“…Now is the time to reflect on how the field developed over the decade and where it is heading. This article approaches it through the lens of the author's contributions 1–46 put in the broader context of the current state of affairs in the field.…”

MLatom 3: A Platform for Machine Learning-Enhanced Computational Chemistry Simulations and Workflows

“…Against this backdrop, the MLatom package started in 2013 as a pure standalone ML package to provide a general-purpose experience for computational chemists akin to the black-box QM packages . The early MLatom could be used for training, testing, and using ML models and their combinations with QM methods (e.g., Δ-learning and learning of Hamiltonian parameters), accurate representation of PES, , sampling of points from data sets, ML-accelerated nonadiabatic dynamics, and materials design . The fast pace of method and software development in QM, MM, ML, and other computational science domains led to MLatom 2, which started to include interfaces to third-party packages .…”

“… 22 The early MLatom could be used for training, testing, and using ML models and their combinations with QM methods (e.g., Δ-learning 23 and learning of Hamiltonian parameters 24 ), accurate representation of PES, 25 , 26 sampling of points from data sets, 26 ML-accelerated nonadiabatic dynamics, 27 and materials design. 28 The fast pace of method and software development in QM, MM, ML, and other computational science domains led to MLatom 2, which started to include interfaces to third-party packages. 29 Such an approach provided a unique opportunity for the package users to choose one of the many established ML models, similar to the users of the traditional QM software who can choose one of the many QM methods.…”

Stable and metastable crystal structures and ammonia dynamics in strontium chloride ammines

“…11,12 Metal halide ammines (MHA) have been extensively studied in the literature due to their potential as cost-effective and safe materials for solid-state ammonia storage. [13][14][15][16][17] One of the most studied members of the metal halide ammine family is magnesium hexammine chloride (Mg(NH 3 ) 6 Cl 2 ) due to its high gravimetric and volumetric ammonia capacity. [18][19][20][21][22] However, Mg(NH 3 ) 6 Cl 2 starts to decompose by releasing the first four NH 3 molecules around 445 K and continues to release the last two NH 3 molecules at 585 and 680 K, respectively.…”