Using a combined theoretical and experimental approach, we investigate the non-adiabatic dynamics of the prototypical ethylene (C(2)H(4)) molecule upon π → π∗ excitation. In this first part of a two part series, we focus on the lifetime of the excited electronic state. The femtosecond time-resolved photoelectron spectrum (TRPES) of ethylene is simulated based on our recent molecular dynamics simulation using the ab initio multiple spawning method with multi-state second order perturbation theory [H. Tao, B. G. Levine, and T. J. Martinez, J. Phys. Chem. A 113, 13656 (2009)]. We find excellent agreement between the TRPES calculation and the photoion signal observed in a pump-probe experiment using femtosecond vacuum ultraviolet (hν = 7.7 eV) pulses for both pump and probe. These results explain the apparent discrepancy over the excited state lifetime between theory and experiment that has existed for ten years, with experiments [e.g., P. Farmanara, V. Stert, and W. Radloff, Chem. Phys. Lett. 288, 518 (1998) and K. Kosma, S. A. Trushin, W. Fuss, and W. E. Schmid, J. Phys. Chem. A 112, 7514 (2008)] reporting much shorter lifetimes than predicted by theory. Investigation of the TRPES indicates that the fast decay of the photoion yield originates from both energetic and electronic factors, with the energetic factor playing a larger role in shaping the signal.

Through a combined experimental and theoretical approach, we study the nonadiabatic dynamics of the prototypical ethylene (C2H4) molecule upon π → π* excitation with 161 nm light. Using a novel experimental apparatus, we combine femtosecond pulses of vacuum ultraviolet and extreme ultraviolet (XUV) radiation with variable delay to perform time resolved photo-ion fragment spectroscopy. In this second part of a two part series, the XUV (17 eV < hν < 23 eV) probe pulses are sufficiently energetic to break the C–C bond in photoionization, or to photoionize the dissociation products of the vibrationally hot ground state. The experimental data is directly compared to excited state ab initio molecular dynamics simulations explicitly accounting for the probe step. Enhancements of the CH2+ and CH3+ photo-ion fragment yields, corresponding to molecules photoionized in ethylene (CH2CH2) and ethylidene (CH3CH) like geometries are observed within 100 fs after π → π* excitation. Quantitative agreement between theory and experiment on the relative CH2+ and CH3+ yields provides experimental confirmation of the theoretical prediction of two distinct conical intersections and their branching ratio [H. Tao, B. G. Levine, and T. J. Martinez, J. Phys. Chem. A. 113, 13656 (2009)]. Evidence for fast, non-statistical, elimination of H2 molecules and H atoms is observed in the time resolved H2+ and H+ signals.

Since the recent advent of deep reinforcement learning for game play [1] and simulated robotic control (e.g. [2]), a multitude of new algorithms have flourished. Most are model-free algorithms which can be categorized into three families: deep Q-learning, policy gradients, and Q-value policy gradients. These have developed along separate lines of research, such that few, if any, code bases incorporate all three kinds. Yet these algorithms share a great depth of common deep reinforcement learning machinery. We are pleased to share rlpyt, which implements all three algorithm families on top of a shared, optimized infrastructure, in a single repository. It contains modular implementations of many common deep RL algorithms in Python using PyTorch [3], a leading deep learning library. rlpyt is designed as a high-throughput code base for small-to medium-scale research in deep RL. This white paper summarizes its features, algorithms implemented, and relation to prior work, and concludes with detailed implementation and usage notes. rlpyt is available at https://github.com/astooke/rlpyt.

In an effort to overcome limitations of reward-driven feature learning in deep reinforcement learning (RL) from images, we propose decoupling representation learning from policy learning. To this end, we introduce a new unsupervised learning (UL) task, called Augmented Temporal Contrast (ATC), which trains a convolutional encoder to associate pairs of observations separated by a short time difference, under image augmentations and using a contrastive loss. In online RL experiments, we show that training the encoder exclusively using ATC matches or outperforms end-to-end RL in most environments. Additionally, we benchmark several leading UL algorithms by pre-training encoders on expert demonstrations and using them, with weights frozen, in RL agents; we find that agents using ATC-trained encoders outperform all others. We also train multi-task encoders on data from multiple environments and show generalization to different downstream RL tasks. Finally, we ablate components of ATC, and introduce a new data augmentation to enable replay of (compressed) latent images from pre-trained encoders when RL requires augmentation. Our experiments span visually diverse RL benchmarks in DeepMind Control, DeepMind Lab, and Atari, and our complete code is available at https://github.com/astooke/rlpyt/tree/ master/rlpyt/ul.