Fully Automated Molecular Design with Atomic Resolution for Desired Thermophysical Properties

Abstract

The value of fine and specialty chemicals is often determined by the specific requirements in their physical and chemical properties. Therefore, it is most desirable to design the structure of chemicals to meet some targeted material properties. In the past, the design of specialty chemicals has been based largely on experience and trial-and-error. However, recent advances in computational chemistry and machine learning can offer a new path to this problem. In this presentation, we demonstrate a successful example where the structure of a chemical of specified value of octanol−water partition coefficient (K ow ) can be predicted by computers. This method consists of two parts, the first being a robust method, the COSMO-SAC activity coefficient model, that predicts the activity coefficient with input of only the molecular structure. The second component of this method is a derivative-free optimization algorithm that searches in the multidimensional structure space for the desired value of K ow . In particular, the genetic algorithm (GA), based on the Darwinian theory of evolution and natural selection, combined with simulated annealing (SA) is adopted for this purpose. Compared to other optimization algorithms, GA can overcome the problem of being trapped in local minima and SA can help improve the convergence. Therefore, the GA−SA combination has been found to be very suitable for molecular design. We show that the value of K ow can be achieved within 1% of the target in 30 generations with a proper set of evolution parameters (including the size of the population, the probability of selection, the rate of temperature annealing, etc.). The same method can be applied to the search for chemicals with other desired properties, such as vapor pressure and solubility.

New Data Structure for Computational Molecular Design with Atomic or Fragment Resolution

“…Furthermore, algorithms are developed to generate a three-dimensional molecular structure (atomic coordinates) from the MDS and MDS from the 3D structure. We note that this MDS has been used earlier for finding the chemicals with desired hydrophobicity and hydrophilicity balance . Here, we show all the detailed internal structure and improved the robustness and reliability of the molecular operation algorithms.…”

“…We note that this MDS has been used earlier for finding the chemicals with desired hydrophobicity and hydrophilicity balance. 61 Here, we show all the detailed internal structure and improved the robustness and reliability of the molecular operation algorithms.…”

“…Besides, we develop algorithms featuring the capability of complicated structure manipulations (e.g., ring formation and opening) and the capability of finding all possible combinations between two fragments. By incorporating this tool with property estimation models and optimization algorithms, the computeraided molecular design 61 for organic compounds and materials by could be achieved.…”

Implementation of Statistical Learning Methods to Develop Guidelines for the Design of PLA-Based Composites with High Tensile Strength Values

“…For instance, machine learning tools are used to predict thermal conductivities of composite materials, ANN models are used to predict Young’s modulus and conductivity of two-phase composites from microstructre properties, and deep learning is used to predict the stiffness of a 3-D material from its microstructure . A very recent application of machine learning methods involves the rather ambitious aim of tailoring chemicals/materials to specific quality ends using generative models trained on the available data, or a general model inversion framework to meet the desired product quality; applications to materials are currently a hot topic. , It should, however, be noted that a majority of the “machine learning meets material science” literature focuses on prediction, or even inverse modeling, mostly using simulated data, while extraction of experience based rules using modern statistical learning tools from historical/observational data is rather scarce. In two of the studies on stainless steels (SS), statistical learning methods were used with this particular aim. , In the former, stress corrosion cracking of 112 samples was modeled using various tools, such as discriminant analysis, nearest neighbors, and decision trees, and decision trees were found to give the highest classification accuracy.…”

“…29 A very recent application of machine learning methods involves the rather ambitious aim of tailoring chemicals/materials to specific quality ends using generative models trained on the available data, 30 or a general model inversion framework to meet the desired product quality; applications to materials are currently a hot topic. 31,32 It should, however, be noted that a majority of the "machine learning meets material science" literature focuses on prediction, or even inverse modeling, mostly using simulated data, while extraction of experience based rules using modern statistical learning tools from historical/ observational data is rather scarce. In two of the studies on stainless steels (SS), statistical learning methods were used with this particular aim.…”