The workshop concludes by tackling a final project: How to generate a CG nebula using volume velocity advection with a simple VOP, where helpful tips about how to create pyro FX advection are provided, as well as a deep look at particle rendering for gaseous FX.
In this work, we propose conditional G-SchNet (cG-SchNet), a conditional generative neural network for the inverse design of molecules. Building on G-SchNet, the model learns conditional distributions depending on structural or chemical properties allowing us to sample corresponding 3d molecular structures. Our architecture is designed to generate molecules of arbitrary size and does not require the specification of a target composition. Consequently, it learns the relationship between the composition of molecules and their physical properties in order to sample candidates exhibiting given target properties, e.g., preferring smaller structures when targeting small polarizabilities. Previously proposed methods have been biased towards one particular set of target property values at a time by adjusting the training objective or data48,51. In contrast, our conditional approach permits searching for molecules with any desired set of target property values after training is completed. It is able to jointly target multiple properties without the need to retrain or otherwise indirectly constrain the sampling process. This provides the foundation for the model to leverage the full information of the training data resulting in increased generalization and data efficiency. We demonstrate that cG-SchNet enables the exploration of sparsely populated regions that are hardly accessible with unconditional models. To this end, we conduct extensive experiments with diverse conditioning targets including chemical properties, atomic compositions and molecular fingerprints. In this way, we generate novel molecules with predefined structural motifs, isomers of a given composition that exhibit specific chemical properties, and novel configurations that jointly optimize HOMO-LUMO gap and energy. This demonstrates that our model enables flexible, guided exploration of chemical compound space.
a Factorization of the conditional joint probability of atom positions and types into a chain of probabilities for placing single atoms one after another. b Results of sampling molecules from target-dependent conditional probability distributions. Distributions of the isotropic polarizability of training structures (orange) and five sets of molecules generated by the same cG-SchNet model (blue curves) conditioned on five different isotropic polarizability target values (color-matching dots above the x-axis). The generated molecule closest to the corresponding target value and not contained in the training data (unseen) is shown above each curve. c Schematic depiction of the atom placement loop. For visualization purposes, we show a planar molecule and a 2d slice of the actual 3d grid distributions in steps 4, 5, and 6.
In many applications, it is advantageous for molecules to possess specific functional groups or structural motifs. These can be correlated with desirable chemical properties, e.g., polar groups that increase solubility, or with improved synthetic accessibility. In order to sample molecules with specific motifs, we condition cG-SchNet on a path-based, 1024 bits long fingerprint that checks molecular graphs for all linear segments of up to seven atoms57 (Supplementary Methods 3). The model is trained on a randomly selected subset of 55k molecules from the QM9 dataset consisting of ~134k organic molecules with up to nine heavy atoms from carbon, nitrogen, oxygen, and fluorine58,59,60. We condition the sampling on fingerprints of unseen molecules, i.e., structures not used during training. Figure 3a shows results for four examples. We observe that the generated molecules have a higher similarity with the target fingerprints than the training data. Furthermore, structures with high target similarity are also sampled with higher probability, as can be seen from the increased similarity score of generated duplicates. In the last column of Fig. 3a, we show sampled molecules with high similarity to each target and see that in each case various structures with perfectly matching fingerprints were found. For reference, we also show the most similar molecule in the training set. Overall, we see that the conditional sampling with cG-SchNet is sensitive to the target fingerprint and allows for the generation of molecules with desired structural motifs. Although there are no molecules with the same fingerprint in the training data for three of the four fingerprint targets, the ML model successfully generates perfectly matching molecules, demonstrating its ability to generalize and explore unseen regions of chemical compound space.
For each training run, 55k reference structures are randomly sampled from the QM9 dataset58,59,60, a collection of 133,885 molecules with up to nine heavy atoms from carbon, nitrogen, oxygen, and fluorine. We removed 915 molecules from the training pool which are deemed invalid by our validation procedure that checks the valency and connectedness of generated structures (see Section Checking validity and uniqueness of generated molecules). For some runs, limited subsets of the training data pool are used, as described in the results (e.g., without C7O2H10 isomers). We train the neural network using 50k randomly sampled molecules and employ the remaining 5k for validation (see Section Neural network training). All molecules shown in figures have been rendered with the 3d visualization package Mayavi66.
Rendering has uses in architecture, video games, simulators, movie and TV visual effects, and design visualization, each employing a different balance of features and techniques. A wide variety of renderers are available for use. Some are integrated into larger modeling and animation packages, some are stand-alone, and some are free open-source projects. On the inside, a renderer is a carefully engineered program based on multiple disciplines, including light physics, visual perception, mathematics, and software development.
Parsons full-residency, two-year MFA program is a STEM-designated, studio based degree with areas of practice including critical design, data visualization, digital fabrication, game design, interaction design, new media art, physical computing, and wearable technology. Through Collaboration Studio courses, students in this 60 credit hour program have the opportunity to work with fellow graduate students in related programs including Communication Design, Data Visualization, and Transdisciplinary Design. 2b1af7f3a8