Materials property prediction for limited datasets enabled by feature selection and joint learning with MODNet

In order to make accurate predictions of material properties, current machine-learning approaches generally require large amounts of data, which are often not available in practice. In this work, MODNet, an all-round framework, is presented which relies on a feedforward neural network, the selection of physically meaningful features, and when applicable, joint-learning. Next to being faster in terms of training time, this approach is shown to outperform current graph-network models on small datasets. In particular, the vibrational entropy at 305 K of crystals is predicted with a mean absolute test error of 0.009 meV/K/atom (four times lower than previous studies). Furthermore, joint learning reduces the test error compared to single-target learning and enables the prediction of multiple properties at once, such as temperature functions. Finally, the selection algorithm highlights the most important features and thus helps to understand the underlying physics.

Vendor-independent software shows limited variability in speckle tracking strain measurements on images of different vendors

Funding Acknowledgements Type of funding sources: None. Background Vendors use proprietary speckle tracking software algorithms for echocardiographic strain measurements, which results in high inter-vendor variability. Little is known about potential advantages or disadvantages of using vendor-independent software in clinical practice. Purpose We therefore investigated the reproducibility, accuracy, and ability to identify scar of strain measurements on images from different vendors by using a vendor-independent software. Methods A vendor-independent software (TomTec Image Arena) was used to analyze datasets of 63 patients which were obtained on four ultrasound machines from different vendors (GE, Philips, Siemens, Toshiba). We measured the tracking feasibility, inter-vendor bias, the relative and absolute test-re-test variability of strain measurements and their ability to detect scar. Cardiac magnetic resonance delayed enhancement images were used as the reference standard of scar definition. Results Tracking feasibility differed depending on the image source (p < 0.05). Variability of global longitudinal strain (GLS) (Figure 1A) was similar (ANOVA p = 0.124) among the images of different vendors whereas variability of segmental longitudinal strain (SLS) (Figure 1B) showed modest difference (ANOVA- peak systolic strain (PS); p = 0.077, end-systolic strain (ES); p = 0.171, post-systolic strain (PSS); p = 0.020). Relative test-re-test variability of GLS showed no differences (ANOVA p = 0.360). Absolute test-re-test errors of SLS measurements showed modest differences among images of different vendors (ANOVA- PS; p = 0.018, ES; p = 0.001, PSS; p = 0.090). No relevant difference in scar detection capability was observed (Figure 1C). Conclusions Vendor independent software leads to low bias among strain measurements on images from different vendors. Likewise, measurement variability and the ability to identify scar becomes similar. Our findings suggest that a vendor independent speckle tracking software could help to overcome inter-vendor bias. To which extend such measurements would be more accurate compared to vendor specific software remains to be determined. Abstract Figure 1

Test–retest variability and reference region-based quantification of 18F-BCPP-EF for imaging mitochondrial complex I in the human brain

Mitochondrial complex I (MC-I) is an essential regulator of brain bioenergetics and can be quantified in the brain using PET radioligand 18F-BCPP-EF. Here we evaluate the test–retest reproducibility of 18F-BCPP-EF in humans, and assess the use of a non-invasive quantification method (standardised uptake value ratio – SUVR). Thirty healthy volunteers had a 90-min dynamic 18F-BCPP-EF scan with arterial blood sampling, five of which received a second scan to be included in the test–retest analysis. Time-activity curves (TAC) were analysed using multilinear analysis 1 (MA1) and the two-tissue compartment model (2TC) to estimate volumes of distribution (VT). Regional SUVR-1 values were calculated from the 70 to 90-min TAC data using the centrum semiovale as a pseudo reference region, and compared to kinetic analysis-derived outcome measures. The mean absolute test–retest variability of VT ranged from 12% to 18% across regions. Both DVR-1and SUVR-1 had improved test–retest variability in the range 2%–7%. SUVR-1 was highly correlated with DVR-1 (r2 = 0.97, n = 30). In conclusion, 18F-BCPP-EF has suitable test–retest reproducibility and can be used to quantify MC-I in clinical studies.

Test-retest reliability of force plate-derived measures of reactive stepping

BackgroundCharacterizing reactive stepping is important to describe the response’s effectiveness. Measures of reactive stepping related to step initiation, execution, and termination phases have been frequently reported to characterize reactive balance control. However, the test-retest reliability of these measures are unknown.Research questionsWhat is the between- and within-session test-retest reliability of various force plate-derived measures of reactive stepping?MethodsNineteen young, healthy adults responded to 6 small (~8-10% of body weight) and 6 large perturbations (~13-15% of body weight) using an anterior lean-and-release system. Tests were conducted on two visits separated by at least two days. Participants were instructed to recover balance in as few steps as possible. Step onset, foot-off, swing, and restabilization times were extracted from force plates. Relative test-retest reliability was determined through intraclass correlation coefficients (ICCs) and 95% confidence intervals (CIs). Absolute test-retest reliability was assessed using the standard error of the measurement (SEM).ResultsFoot-off and swing times had the highest between- and within-session test-retest reliabilities regardless of perturbation size (between-session ICC=0.898–0.942; within-session ICC=0.455–0.753). Conversely, step onset and restabilization time had lower ICCs and wider CIs (between-session ICC=0.495–0.825; within-session ICC=−0.040–0.174). Between-session test-retest reliability was higher (ICC=0.495-0.942) for all measures than within-session test-retest reliability (ICC=−0.040–0.753). SEMs were low (3–10% of mean) for all measures, except time to restabilization (SEM=15-20% of mean), indicating good absolute reliability.SignificanceThese findings suggest multiple baseline sessions are needed for measuring restabilization and step onset times. The SEMs provide an index for measuring meaningful change due to an intervention.

Predicting Intermetallic Surface Energies with High-Throughput DFT and Convolutional Neural Networks

Surface energy of inorganic crystals is crucial in understanding experimentally-relevant surface properties and thus important in designing materials for many applications including catalysis. Predictive methods and datasets exist for surface energies of monometallic crystals but predicting these properties for bimetallic or more complicated surfaces is an open challenge. Here we present a workflow for predicting surface energies \textit{ab initio} using high-throughput DFT and a machine learning framework. We calculate the surface energy of 3,285 intermetallic alloys with combinations of 36 elements and 47 space groups. We used this high-throughput workflow to seed a database of surface energies, which we used to train a crystal graph convolutional neural network (CGCNN). The CGCNN model was able to predict surface energies with a mean absolute test error of 0.0082 eV/angstrom^2 and can qualitatively reproduce nanoparticle surface distributions (Wulff constructions). Our workflow provides quantitative insights into which surfaces are more stable and therefore more realistic. It allows us to down-select interesting candidates that we can study with robust theoretical and experimental methods for applications such as catalysts screening and nanomaterials synthesis.

Predicting Intermetallic Surface Energies with High-Throughput DFT and Convolutional Neural Networks

Surface energy of inorganic crystals is crucial in understanding experimentally-relevant surface properties and thus important in designing materials for many applications including catalysis. Predictive methods and datasets exist for surface energies of monometallic crystals but predicting these properties for bimetallic or more complicated surfaces is an open challenge. Here we present a workflow for predicting surface energies \textit{ab initio} using high-throughput DFT and a machine learning framework. We calculate the surface energy of 3,285 intermetallic alloys with combinations of 36 elements and 47 space groups. We used this high-throughput workflow to seed a database of surface energies, which we used to train a crystal graph convolutional neural network (CGCNN). The CGCNN model was able to predict surface energies with a mean absolute test error of 0.0082 eV/angstrom^2 and can qualitatively reproduce nanoparticle surface distributions (Wulff constructions). Our workflow provides quantitative insights into which surfaces are more stable and therefore more realistic. It allows us to down-select interesting candidates that we can study with robust theoretical and experimental methods for applications such as catalysts screening and nanomaterials synthesis.

Disrupting Education? Experimental Evidence on Technology-Aided Instruction in India

We study the impact of a personalized technology-aided after-school instruction program in middle-school grades in urban India using a lottery that provided winners with free access to the program. Lottery winners scored 0.37 σ higher in math and 0.23 σ higher in Hindi over just a 4.5-month period. IV estimates suggest that attending the program for 90 days would increase math and Hindi test scores by 0.6 σ and 0.39 σ respectively. We find similar absolute test score gains for all students, but much greater relative gains for academically-weaker students. Our results suggest that well-designed, technology-aided instruction programs can sharply improve productivity in delivering education. (JEL I21, I26, I28, J24, O15)

Strength and Interface Adhesion Properties of in-Plane Shear Loaded Thick Adhesive Joint

This paper addresses the quality of the interface bonded joints in layers of timber elements. The shear performance was studied to assess the suitability of adhesive to bond timber with different thickness of glueline and timber densities. Since there is absolute test method in establishing the shear strength of the surface bonds between layers timber elements, two test methods were used namely Thick Adherend Shear Test [TAST] and Lap Shear Test. The adhesived used is Sikadur-30 and timber used are Sesendok, Bintangor and Kempas with average densities 400-600kg/m3, 600-800kg/m3 and 800-1000kg/m3 respectively. Obtained results suggest that the interface stress distribution are related to the thicknesses of adhesive and densities of timber.

Homogenization and Trend Analysis of the 1958–2016 In Situ Surface Solar Radiation Records in China

This paper presents a method to homogenize China’s surface solar radiation (SSR) data and uses the resulting homogenized SSR data to assess the SSR trend over the period 1958–2016. Neighboring surface sunshine duration (SSD) data are used as reference data to assess the SSR data homogeneity. A principal component analysis is applied to build a reference series, which is proven to be less sensitive to occasional data issues than using the arithmetic mean of data from adjacent stations. A relative or absolute test is applied to detect changepoints, depending on whether or not a suitable reference series is available. A quantile-matching method is used to adjust the data to diminish the inhomogeneities. As a result, 60 out of the 119 SSR stations were found to have inhomogeneity issues. These were mainly caused by changes in instrument and observation schedule. The nonclimatic changes exaggerated the SSR change rates in 1991–93 and resulted in a sudden rise in the national average SSR series, causing an unrealistically drastic trend reversal in the 1990s. This was diminished by the data homogenization. The homogenized data show that the national average SSR has been declining significantly over the period 1958–90; this dimming trend mostly diminished over the period 1991–2005 and was replaced by a brightening trend in the recent decade. From the homogenized SSR data, the 1958–90 and 1958–2005 dimming rate is estimated to be −6.13 ± 0.47 and −5.08 ± 0.27 W m−2 decade−1, respectively, and the 2005–16 brightening rate is 6.13 ± 1.77 W m−2 decade−1.