Authors' response to short comment by Lutz Gross

Abstract

The spatiotemporal distribution and characterization of aerosol particles are usually determined by remotesensing and optical in situ measurements. These measurements are indirect with respect to microphysical properties, and thus inversion techniques are required to determine the aerosol microphysics. Scattering theory provides the link between microphysical and optical properties; it is not only needed for such inversions but also for radiative budget calculations and climate modeling. However, optical modeling can be very time-consuming, in particular if nonspherical particles or complex ensembles are involved.In this paper we present the MOPSMAP package (Modeled optical properties of ensembles of aerosol particles), which is computationally fast for optical modeling even in the case of complex aerosols. The package consists of a data set of pre-calculated optical properties of single aerosol particles, a Fortran program to calculate the properties of userdefined aerosol ensembles, and a user-friendly web interface for online calculations. Spheres, spheroids, and a small set of irregular particle shapes are considered over a wide range of sizes and refractive indices. MOPSMAP provides the fundamental optical properties assuming random particle orientation, including the scattering matrix for the selected wavelengths. Moreover, the output includes tables of frequently used properties such as the single-scattering albedo, the asymmetry parameter, or the lidar ratio. To demonstrate the wide range of possible MOPSMAP applications, a selection of examples is presented, e.g., dealing with hygroscopic growth, mixtures of absorbing and non-absorbing particles, the relevance of the size equivalence in the case of nonspherical particles, and the variability in volcanic ash microphysics.The web interface is designed to be intuitive for expert and nonexpert users. To support users a large set of default settings is available, e.g., several wavelength-dependent refractive indices, climatologically representative size distributions, and a parameterization of hygroscopic growth. Calculations are possible for single wavelengths or user-defined sets (e.g., of specific remote-sensing application). For expert users more options for the microphysics are available. Plots for immediate visualization of the results are shown. The complete output can be downloaded for further applications. All input parameters and results are stored in the user's personal folder so that calculations can easily be reproduced. The web interface is provided at https://mopsmap.net (last access: 9 July 2018) and the Fortran program including the data set is freely available for offline calculations, e.g., when large numbers of different runs for sensitivity studies are to be made.

The Open Catalyst 2022 (OC22) Dataset and Challenges for Oxide Electrocatalysts

“…Unlike descriptor based models, − where hand crafted representations are used to describe atomic environments, GNNs learn atomic representations through several message passing steps . Consistent with related work, ,, graphs are constructed with atoms treated as nodes and interactions between atoms as edges. Periodic boundary conditions are accounted for in graph construction consistent with OC20.…”

“…Initial baseline models like CGCNN and SchNet focused on local environment representations. Key advances since then include invariant angular interactions (DimeNet/DimeNet++ , ), faster and more accurate but nonenergy conserving models (ForceNet and SpinConv), and triple/quadruplet interactions (GemNet-dT, GemNet-XL, and GemNet-OC). Other approaches include the use of transformers (3D-Graphormer) and more effective augmentation and learning strategies (Noisy-Nodes).…”

Machine learning in energy chemistry: introduction, challenges and perspectives

“…Also, GCNN has been applied to conquer the limitation of traditional methods, which do not consider the spatial information. In Gasteiger et al, 200 researchers construct the directional message passing neural network (DimeNet), which embed the messages passed between atoms by considering In conclusion, machine learning has been widely applied to find the PESs and atomic forces. With the development of machine learning algorithms, different descriptors and regression methods are applied and obtain a great progress.…”

“…Also, GCNN has been applied to overcome the limitations of traditional methods, which do not consider the spatial information. Gasteiger et al 200 constructed a directional message passing neural network (DimeNet) that embeds the messages passed between atoms by considering directional information.…”

Augmenting Polymer Datasets by Iterative Rearrangement

“…Notable examples include randomized SMILES , and crystal structure augmentation . Unlike visual data which can be augmented in many ways, generating novel molecular representations is challenging because molecular structures are naturally represented as graphs, which are rotationally, translationally, and permutationally invariant. − GNNs, an increasingly popular method for representing polymers, , could potentially benefit from the new substructural information generated by augmentation via iterative rearrangement. Another alternative for improving polymer property prediction is the pre-training of DL models on large unlabeled data to learn the syntax of SMILES and meaning captured by SMILES which can be later fine-tuned on small polymer datasets .…”