skip to main content

Title: Anharmonic contribution to the stabilization of Mg(OH) 2 from first principles
Geometrical and vibrational characterization of magnesium hydroxide was performed using density functional theory. Four possible crystal symmetries were explored: P 3̄ (No. 147, point group −3), C 2/ m (No. 12, point group 2), P 3 m 1 (No. 156, point group 3 m ) and P 3̄ m 1 (No. 164, point group −3 m ) which are the currently accepted geometries found in the literature. While a lot of work has been performed on Mg(OH) 2 , in particular for the P 3̄ m 1 phase, there is still a debate on the observed ground state crystal structure and the anharmonic effects of the OH vibrations on the stabilization of the crystal structure. In particular, the stable positions of hydrogen are not yet defined precisely, which have implications in the crystal symmetry, the vibrational excitations, and the thermal stability. Previous work has assigned the P 3̄ m 1 polymorph as the low energy phase, but it has also proposed that hydrogens are disordered and they could move from their symmetric position in the P 3̄ m 1 structure towards P 3̄. In this paper, we examine the stability of the proposed phases by using different descriptors. We compare more » the XRD patterns with reported experimental results, and a fair agreement is found. While harmonic vibrational analysis shows that most phases have imaginary modes at 0 K, anharmonic vibrational analysis indicates that at room temperature only the C 2/ m phase is stabilized, whereas at higher temperatures, other phases become thermally competitive. « less
; ; ; ; ; ; ;
Award ID(s):
1740111 1434897
Publication Date:
Journal Name:
Physical Chemistry Chemical Physics
Page Range or eLocation-ID:
17799 to 17808
Sponsoring Org:
National Science Foundation
More Like this
  1. FeAs 2−x Se x ( x = 0.30–1.0) samples were synthesized as phase pure powders by conventional solid-state techniques and as single crystals ( x = 0.50) from chemical vapor transport. The composition of the crystals was determined to be Fe 1.025(3) As 1.55(3) Se 0.42(3) , crystallizing in the marcasite structure type, Pnnm space group. FeAs 2−x Se x (0 < x < 1) was found to undergo a marcasite-to-arsenopyrite ( P 2 1 / c space group) structural phase transition at x ∼ 0.65. The structures are similar, with the marcasite structure best described as a solid solutionmore »of As/Se, whereas the arsenopyrite has ordered anion sites. Magnetic susceptibility and thermoelectric property measurements from 300–2 K were performed on single crystals, FeAs 1.50 Se 0.50 . Paramagnetic behavior is observed from 300 to 17 K and a Seebeck coefficient of −33 μV K −1 , an electrical resistivity of 4.07 mΩ cm, and a very low κ l of 0.22 W m −1 K −1 at 300 K are observed. In order to determine the impact of the structural transition on the high-temperature thermoelectric properties, polycrystalline FeAs 2−x Se x ( x = 0.30, 0.75, 0.85, 1.0) samples were consolidated into dense pellets for measurements of thermoelectric properties. The x = 0.85 sample shows the best thermoelectric performance. The electronic structure of FeAsSe was calculated with DFT and transport properties were approximately modeled above 500 K.« less
  2. Obeid, Iyad Selesnick (Ed.)
    Electroencephalography (EEG) is a popular clinical monitoring tool used for diagnosing brain-related disorders such as epilepsy [1]. As monitoring EEGs in a critical-care setting is an expensive and tedious task, there is a great interest in developing real-time EEG monitoring tools to improve patient care quality and efficiency [2]. However, clinicians require automatic seizure detection tools that provide decisions with at least 75% sensitivity and less than 1 false alarm (FA) per 24 hours [3]. Some commercial tools recently claim to reach such performance levels, including the Olympic Brainz Monitor [4] and Persyst 14 [5]. In this abstract, we describemore »our efforts to transform a high-performance offline seizure detection system [3] into a low latency real-time or online seizure detection system. An overview of the system is shown in Figure 1. The main difference between an online versus offline system is that an online system should always be causal and has minimum latency which is often defined by domain experts. The offline system, shown in Figure 2, uses two phases of deep learning models with postprocessing [3]. The channel-based long short term memory (LSTM) model (Phase 1 or P1) processes linear frequency cepstral coefficients (LFCC) [6] features from each EEG channel separately. We use the hypotheses generated by the P1 model and create additional features that carry information about the detected events and their confidence. The P2 model uses these additional features and the LFCC features to learn the temporal and spatial aspects of the EEG signals using a hybrid convolutional neural network (CNN) and LSTM model. Finally, Phase 3 aggregates the results from both P1 and P2 before applying a final postprocessing step. The online system implements Phase 1 by taking advantage of the Linux piping mechanism, multithreading techniques, and multi-core processors. To convert Phase 1 into an online system, we divide the system into five major modules: signal preprocessor, feature extractor, event decoder, postprocessor, and visualizer. The system reads 0.1-second frames from each EEG channel and sends them to the feature extractor and the visualizer. The feature extractor generates LFCC features in real time from the streaming EEG signal. Next, the system computes seizure and background probabilities using a channel-based LSTM model and applies a postprocessor to aggregate the detected events across channels. The system then displays the EEG signal and the decisions simultaneously using a visualization module. The online system uses C++, Python, TensorFlow, and PyQtGraph in its implementation. The online system accepts streamed EEG data sampled at 250 Hz as input. The system begins processing the EEG signal by applying a TCP montage [8]. Depending on the type of the montage, the EEG signal can have either 22 or 20 channels. To enable the online operation, we send 0.1-second (25 samples) length frames from each channel of the streamed EEG signal to the feature extractor and the visualizer. Feature extraction is performed sequentially on each channel. The signal preprocessor writes the sample frames into two streams to facilitate these modules. In the first stream, the feature extractor receives the signals using stdin. In parallel, as a second stream, the visualizer shares a user-defined file with the signal preprocessor. This user-defined file holds raw signal information as a buffer for the visualizer. The signal preprocessor writes into the file while the visualizer reads from it. Reading and writing into the same file poses a challenge. The visualizer can start reading while the signal preprocessor is writing into it. To resolve this issue, we utilize a file locking mechanism in the signal preprocessor and visualizer. Each of the processes temporarily locks the file, performs its operation, releases the lock, and tries to obtain the lock after a waiting period. The file locking mechanism ensures that only one process can access the file by prohibiting other processes from reading or writing while one process is modifying the file [9]. The feature extractor uses circular buffers to save 0.3 seconds or 75 samples from each channel for extracting 0.2-second or 50-sample long center-aligned windows. The module generates 8 absolute LFCC features where the zeroth cepstral coefficient is replaced by a temporal domain energy term. For extracting the rest of the features, three pipelines are used. The differential energy feature is calculated in a 0.9-second absolute feature window with a frame size of 0.1 seconds. The difference between the maximum and minimum temporal energy terms is calculated in this range. Then, the first derivative or the delta features are calculated using another 0.9-second window. Finally, the second derivative or delta-delta features are calculated using a 0.3-second window [6]. The differential energy for the delta-delta features is not included. In total, we extract 26 features from the raw sample windows which add 1.1 seconds of delay to the system. We used the Temple University Hospital Seizure Database (TUSZ) v1.2.1 for developing the online system [10]. The statistics for this dataset are shown in Table 1. A channel-based LSTM model was trained using the features derived from the train set using the online feature extractor module. A window-based normalization technique was applied to those features. In the offline model, we scale features by normalizing using the maximum absolute value of a channel [11] before applying a sliding window approach. Since the online system has access to a limited amount of data, we normalize based on the observed window. The model uses the feature vectors with a frame size of 1 second and a window size of 7 seconds. We evaluated the model using the offline P1 postprocessor to determine the efficacy of the delayed features and the window-based normalization technique. As shown by the results of experiments 1 and 4 in Table 2, these changes give us a comparable performance to the offline model. The online event decoder module utilizes this trained model for computing probabilities for the seizure and background classes. These posteriors are then postprocessed to remove spurious detections. The online postprocessor receives and saves 8 seconds of class posteriors in a buffer for further processing. It applies multiple heuristic filters (e.g., probability threshold) to make an overall decision by combining events across the channels. These filters evaluate the average confidence, the duration of a seizure, and the channels where the seizures were observed. The postprocessor delivers the label and confidence to the visualizer. The visualizer starts to display the signal as soon as it gets access to the signal file, as shown in Figure 1 using the “Signal File” and “Visualizer” blocks. Once the visualizer receives the label and confidence for the latest epoch from the postprocessor, it overlays the decision and color codes that epoch. The visualizer uses red for seizure with the label SEIZ and green for the background class with the label BCKG. Once the streaming finishes, the system saves three files: a signal file in which the sample frames are saved in the order they were streamed, a time segmented event (TSE) file with the overall decisions and confidences, and a hypotheses (HYP) file that saves the label and confidence for each epoch. The user can plot the signal and decisions using the signal and HYP files with only the visualizer by enabling appropriate options. For comparing the performance of different stages of development, we used the test set of TUSZ v1.2.1 database. It contains 1015 EEG records of varying duration. The any-overlap performance [12] of the overall system shown in Figure 2 is 40.29% sensitivity with 5.77 FAs per 24 hours. For comparison, the previous state-of-the-art model developed on this database performed at 30.71% sensitivity with 6.77 FAs per 24 hours [3]. The individual performances of the deep learning phases are as follows: Phase 1’s (P1) performance is 39.46% sensitivity and 11.62 FAs per 24 hours, and Phase 2 detects seizures with 41.16% sensitivity and 11.69 FAs per 24 hours. We trained an LSTM model with the delayed features and the window-based normalization technique for developing the online system. Using the offline decoder and postprocessor, the model performed at 36.23% sensitivity with 9.52 FAs per 24 hours. The trained model was then evaluated with the online modules. The current performance of the overall online system is 45.80% sensitivity with 28.14 FAs per 24 hours. Table 2 summarizes the performances of these systems. The performance of the online system deviates from the offline P1 model because the online postprocessor fails to combine the events as the seizure probability fluctuates during an event. The modules in the online system add a total of 11.1 seconds of delay for processing each second of the data, as shown in Figure 3. In practice, we also count the time for loading the model and starting the visualizer block. When we consider these facts, the system consumes 15 seconds to display the first hypothesis. The system detects seizure onsets with an average latency of 15 seconds. Implementing an automatic seizure detection model in real time is not trivial. We used a variety of techniques such as the file locking mechanism, multithreading, circular buffers, real-time event decoding, and signal-decision plotting to realize the system. A video demonstrating the system is available at: The final conference submission will include a more detailed analysis of the online performance of each module. ACKNOWLEDGMENTS Research reported in this publication was most recently supported by the National Science Foundation Partnership for Innovation award number IIP-1827565 and the Pennsylvania Commonwealth Universal Research Enhancement Program (PA CURE). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the official views of any of these organizations. REFERENCES [1] A. Craik, Y. He, and J. L. Contreras-Vidal, “Deep learning for electroencephalogram (EEG) classification tasks: a review,” J. Neural Eng., vol. 16, no. 3, p. 031001, 2019. [2] A. C. Bridi, T. Q. Louro, and R. C. L. Da Silva, “Clinical Alarms in intensive care: implications of alarm fatigue for the safety of patients,” Rev. Lat. Am. Enfermagem, vol. 22, no. 6, p. 1034, 2014. [3] M. Golmohammadi, V. Shah, I. Obeid, and J. Picone, “Deep Learning Approaches for Automatic Seizure Detection from Scalp Electroencephalograms,” in Signal Processing in Medicine and Biology: Emerging Trends in Research and Applications, 1st ed., I. Obeid, I. Selesnick, and J. Picone, Eds. New York, New York, USA: Springer, 2020, pp. 233–274. [4] “CFM Olympic Brainz Monitor.” [Online]. Available: [Accessed: 17-Jul-2020]. [5] M. L. Scheuer, S. B. Wilson, A. Antony, G. Ghearing, A. Urban, and A. I. Bagic, “Seizure Detection: Interreader Agreement and Detection Algorithm Assessments Using a Large Dataset,” J. Clin. Neurophysiol., 2020. [6] A. Harati, M. Golmohammadi, S. Lopez, I. Obeid, and J. Picone, “Improved EEG Event Classification Using Differential Energy,” in Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium, 2015, pp. 1–4. [7] V. Shah, C. Campbell, I. Obeid, and J. Picone, “Improved Spatio-Temporal Modeling in Automated Seizure Detection using Channel-Dependent Posteriors,” Neurocomputing, 2021. [8] W. Tatum, A. Husain, S. Benbadis, and P. Kaplan, Handbook of EEG Interpretation. New York City, New York, USA: Demos Medical Publishing, 2007. [9] D. P. Bovet and C. Marco, Understanding the Linux Kernel, 3rd ed. O’Reilly Media, Inc., 2005. [10] V. Shah et al., “The Temple University Hospital Seizure Detection Corpus,” Front. Neuroinform., vol. 12, pp. 1–6, 2018. [11] F. Pedregosa et al., “Scikit-learn: Machine Learning in Python,” J. Mach. Learn. Res., vol. 12, pp. 2825–2830, 2011. [12] J. Gotman, D. Flanagan, J. Zhang, and B. Rosenblatt, “Automatic seizure detection in the newborn: Methods and initial evaluation,” Electroencephalogr. Clin. Neurophysiol., vol. 103, no. 3, pp. 356–362, 1997.« less
  3. The hydrogen-bonding environments at the COOH moiety in eight polycrystalline polymorphs of palmitic acid are explored using solid-state NMR. Although most phases have no previously reported crystal structure, measured 13 C chemical shift tensors for COOH moieties, combined with DFT modeling establish that all phases crystallize with a cyclic dimer ( R 22(8)) hydrogen bonding arrangement. Phases A 2 , B m and E m have localized OH hydrogens while phase C has a dynamically disordered OH hydrogen. The phase designated A s is a mix of five forms, including 27.4% of B m and four novel phases not fullymore »characterized here due to insufficient sample mass. For phases A 2 , B m , E m , and C the anisotropic uncertainties in the COOH hydrogen atom positions are established using a Monte Carlo sampling scheme. Sampled points are retained or rejected at the ±1 σ level based upon agreement of DFT computed 13 COOH tensors with experimental values. The collection of retained hydrogen positions bear a remarkable resemblance to the anisotropic displacement parameters ( i.e. thermal ellipsoids) from diffraction studies. We posit that this similarity is no mere coincidence and that the two are fundamentally related. The volumes of NMR-derived anisotropic displacement ellipsoids for phases with localized OH hydrogens are 4.1 times smaller than those derived from single crystal X-ray diffraction and 1.8 times smaller than the volume of benchmark single crystal neutron diffraction values.« less
  4. Understanding H 2 binding and activation is important in the context of designing transition metal catalysts for many processes, including hydrogenation and the interconversion of H 2 with protons and electrons. This work reports the first thermodynamic and kinetic H 2 binding studies for an isostructural series of first-row metal complexes: NiML, where M = Al ( 1 ), Ga ( 2 ), and In ( 3 ), and L = [N( o -(NCH 2 P i Pr 2 )C 6 H 4 ) 3 ] 3− . Thermodynamic free energies (Δ G °) and free energies of activation (Δmore »G ‡ ) for binding equilibria were obtained via variable-temperature 31 P NMR studies and lineshape analysis. The supporting metal exerts a large influence on the thermodynamic favorability of both H 2 and N 2 binding to Ni, with Δ G ° values for H 2 binding found to span nearly the entire range of previous reports. The non-classical H 2 adduct, (η 2 -H 2 )NiInL ( 3 -H 2 ), was structurally characterized by single-crystal neutron diffraction—the first such study for a Ni(η 2 -H 2 ) complex or any d 10 M(η 2 -H 2 ) complex. UV-Vis studies and TD-DFT calculations identified specific electronic structure perturbations of the supporting metal which poise NiML complexes for small-molecule binding. ETS-NOCV calculations indicate that H 2 binding primarily occurs via H–H σ-donation to the Ni 4p z -based LUMO, which is proposed to become energetically accessible as the Ni(0)→M( iii ) dative interaction increases for the larger M( iii ) ions. Linear free-energy relationships are discussed, with the activation barrier for H 2 binding (Δ G ‡ ) found to decrease proportionally for more thermodynamically favorable equilibria. The Δ G ° values for H 2 and N 2 binding to NiML complexes were also found to be more exergonic for the larger M( iii ) ions.« less
  5. The high-pressure structure and stability of the calcic amphibole tremolite (Ca2Mg5Si8O22(OH)2) was investigated to ~40 GPa at 300 K by single-crystal X-ray diffraction using synchrotron radiation. C2/m symmetry tremolite displays a broader metastability range than previously studied clinoamphiboles, exhibiting no first-order phase transition up to 40 GPa. Axial parameter ratios a/b and a/c, in conjunction with finite strain versus normalized pressure trends, indicate that changes in compressional behavior occur at pressures of ~5 and ~20 GPa. An analysis of the finite strain trends, using third-order Birch-Murnaghan equations of state, resulted in bulk moduli (𝐾) of 72(7), 77(2), and 61(1) GPamore »for the compressional regimes from 0-5 GPa (regime I), 5-20 GPa (II), and above 20 GPa (III), respectively, and accompanying pressure-derivatives of the bulk moduli (𝐾′) of 8.6(42), 6.0(3), and 10.0(2). The results are consistent with first-principle theoretical calculations of tremolite elasticity. The axial compressibility ratios of tremolite, determined as 𝛽a : 𝛽b : 𝛽c = 2.22:1.0:0.78 (regime I), 2.12:1.0:0.96 (II), and 1.03:1.0:0.75 (III), demonstrate a substantial reduction of the compressional anisotropy of tremolite at high pressures, which is a notable contrast with the increasingly anisotropic compressibility observed in the high-pressure polymorphs of the clinoamphibole grunerite. The shift in compression-regime at 5 GPa (I-II) transition is ascribed to stiffening along the crystallographic a-axis corresponding to closure of the vacant A-site in the structure, and a shift in the topology of the a-oriented surfaces of the structural I-beam from concave to convex. The II-III regime shift at 20 GPa corresponds to an increasing rate of compaction of the Ca-polyhedra and increased distortion of the Mg-octahedral sites, processes which dictate compaction in both high-pressure compression-regimes. Bond-valence analyses of the tremolite structure under pressure show dramatic overbonding of the Ca-cations (75% at 30 GPa), with significant Mg-cation overbonding as well (40%). These imply that tremolite’s notable metastability range hinges on the calcium cation’s bonding environment. The 8-fold coordinated Ca-polyhedron accommodates significant compaction under pressure, while the geometry of the Ca-O polyhedron becomes increasingly regular and inhibits the reorientation of the tetrahedral chains that generate phase transitions observed in other clinoamphiboles. Peak/background ratio of diffraction data collected above 40 GPa and our equation of state determination of bulk moduli and compressibilities of tremolite in regime III, in concert with the results of our previous Raman study, suggest that C2/m tremolite may be approaching the limit of its metastability above 40 GPa. Our results have relevance for both the metastable compaction of tremolite during impact events, and for possible metastable persistence of tremolite within cold subduction zones within the Earth.« less