Search

A method is introduced to determine lattice parameters using machine learning. Analysis is presented of the impact of experimental conditions on machine learning prediction, and possibilities for automated unit-cell solution are explored., A key step in the analysis of powder X-ray diffraction (PXRD) data is the accurate determination of unit-cell lattice parameters. This step often requires significant human intervention and is a bottleneck that hinders efforts towards automated analysis. This work develops a series of one-dimensional convolutional neural networks (1D-CNNs) trained to provide lattice parameter estimates for each crystal system. A mean absolute percentage error of approximately 10% is achieved for each crystal system, which corresponds to a 100- to 1000-fold reduction in lattice parameter search space volume. The models learn from nearly one million crystal structures contained within the Inorganic Crystal Structure Database and the Cambridge Structural Database and, due to the nature of these two complimentary databases, the models generalize well across chemistries. A key component of this work is a systematic analysis of the effect of different realistic experimental non-idealities on model performance. It is found that the addition of impurity phases, baseline noise and peak broadening present the greatest challenges to learning, while zero-offset error and random intensity modulations have little effect. However, appropriate data modification schemes can be used to bolster model performance and yield reasonable predictions, even for data which simulate realistic experimental non-idealities. In order to obtain accurate results, a new approach is introduced which uses the initial machine learning estimates with existing iterative whole-pattern refinement schemes to tackle automated unit-cell solution.

The absence of significant detectable signatures of organic molecules in the atmosphere and on the surface of Mars is a major unsolved puzzle. One possible explanation is that perchlorate-rich Martian soils, activated by solar ultraviolet (UV) radiation, create an environment favorable for the rapid oxidation of organics such as alkanes (including methane or CH₄). In this paper, we measured product formation rates from the methane-perchlorate-UV system at room temperature. Our results show that magnesium perchlorate (Mg(ClO₄)₂•6H₂O) surfaces exposed to UV light at wavelengths reaching the Mars' surface accelerate the decomposition of methane (CH₄), resulting in the formation of carbon dioxide (CO₂), carbon monoxide (CO), and volatile chlorine oxides. The production rates for CO₂ and CO on UV-activated perchlorate surfaces are accelerated by a factor of 2.5 and 4.5, respectively, compared to those in the absence of perchlorate. In addition, with UV radiation exposure, perchlorate (ClO₄⁻) decomposes to chlorate (ClO₃⁻) and chlorine oxides. These results are incorporated into a simple box model to estimate the near-surface atmospheric methane lifetime. The model gives a lower bound of the lifetime on the order of hours to days, substantially shorter than ~300 yrs. calculated from methane loss by gas-phase chemistry alone.

We investigate the structure of the loss function landscape for neural networks subject to dataset mislabelling, increased training set diversity, and reduced node connectivity, using various techniques developed for energy landscape exploration. The benchmarking models are classification problems for atomic geometry optimisation and hand-written digit prediction. We consider the effect of varying the size of the atomic configuration space used to generate initial geometries and find that the number of stationary points increases rapidly with the size of the training configuration space. We introduce a measure of node locality to limit network connectivity and perturb permutational weight symmetry, and examine how this parameter affects the resulting landscapes. We find that highly-reduced systems have low capacity and exhibit landscapes with very few minima. On the other hand, small amounts of reduced connectivity can enhance network expressibility and can yield more complex landscapes. Investigating the effect of deliberate classification errors in the training data, we find that the variance in testing AUC, computed over a sample of minima, grows significantly with the training error, providing new insight into the role of the variance-bias trade-off when training under noise. Finally, we illustrate how the number of local minima for networks with two and three hidden layers, but a comparable number of variable edge weights, increases significantly with the number of layers, and as the number of training data decreases. This work helps shed further light on neural network loss landscapes and provides guidance for future work on neural network training and optimisation., epsrc

© 2017 SPIE. We consider the problem of identifying frames in a cardiac ultrasound video associated with left ventricular chamber end-systolic (ES, contraction) and end-diastolic (ED, expansion) phases of the cardiac cycle. Our procedure involves a simple application of non-negative matrix factorization (NMF) to a series of frames of a video from a single patient. Rank-2 NMF is performed to compute two end-members. The end members are shown to be close representations of the actual heart morphology at the end of each phase of the heart function. Moreover, the entire time series can be represented as a linear combination of these two end-member states thus providing a very low dimensional representation of the time dynamics of the heart. Unlike previous work, our methods do not require any electrocardiogram (ECG) information in order to select the end-diastolic frame. Results are presented for a data set of 99 patients including both healthy and diseased examples.