Search

A dataset is the basis of deep learning model development, and the success of deep learning models heavily relies on the quality and size of the dataset. In this work, we present a new data preparation protocol and build a large fragment-based dataset Frag20, which consists of optimized 3D geometries and calculated molecular properties from Merck molecular force field (MMFF) and DFT at the B3LYP/6-31G* level of theory for more than half a million molecules composed of H, B, C, O, N, F, P, S, Cl, and Br with no larger than 20 heavy atoms. Based on the new dataset, we develop robust molecular energy prediction models using a simplified PhysNet architecture for both DFT-optimized and MMFF-optimized geometries, which achieve better than or close to chemical accuracy (1 kcal/mol) on multiple test sets, including CSD20 and Plati20 based on experimental crystal structures.

Efficient synchronization is of great importance for fiber optical coherent communication system. Here, we propose an accurate and low-complexity timing/frequency synchronization scheme based on sparse fast Fourier transform (S-FFT). The proposed scheme consists of coarse timing/frequency synchronization, fine timing synchronization, and fine frequency synchronization. Inspired by the idea of S-FFT, we take full advantages of the sparse nature of training symbols (TSs) and realize the synchronization in the frequency domain. The proposed scheme enables accurate timing synchronization at low signal-to-noise ratio (SNR) region. A complete frequency offset estimation (FOE) range of [−symbol rate/2, + symbol rate/2] can be realized with high resolution. Finally, the proposed scheme is experimentally verified through 10 Gbaud dual-polarization (DP) 16/32-quadrature amplitude modulation (QAM) transmission, without the BER performance penalty. In particular, the total computation complexity is reduced by nearly 200 times in comparison with that using conventional sliding window correlation scheme.

The use of neural networks to predict molecular properties calculated from high level quantum mechanical calculations has made significant advances in recent years, but most models need input geometries from DFT optimizations which limit their applicability in practice. In this work, we explored how machine learning can be used to predict molecular atomization energies and conformation stability using optimized geometries from Merck Molecular Force Field (MMFF). Based on the recently introduced deep tensor neural network (DTNN) approach, we first improved its training efficiency and performed an extensive search of its hyperparameters, and developed a DTNN_7ib model which has a test accuracy of 0.34 kcal/mol mean absolute error (MAE) on QM9 dataset. Then using atomic vector representations in the DTNN_7ib model, we employed transfer learning (TL) strategy to train readout layers on the QM9(M) dataset, in which QM properties are the same as in QM9 [calculated at the B3LYP/6–31G(2df,p) level] while molecular geometries are corresponding local minima optimized with MMFF94 force field. The developed TL_QM9(M) model can achieve an MAE of 0.79 kcal/mol using MMFF optimized geometries. Furthermore, we demonstrated that the same transfer learning strategy with the same atomic vector representation can be used to develop a machine learning model that can achieve an MAE of 0.51 kcal/mol in molecular energy prediction using MMFF geometries for an eMol9_C(M) conformation dataset, which consists of 9959 molecules and 88,234 conformations with energies calculated at the B3LYP/6–31G* level. Our results indicate that DFT-level accuracy of molecular energy prediction can be achieved using force-field optimized geometries and atomic vector representations learned from deep tensor neural network, and integrated molecular modeling and machine learning would be a promising approach to develop more powerful computational tools for molecular conformation analysis.

We discuss various applications of machine learning techniques in different aspects of optical communications and networking including optical performance monitoring, fiber nonlinearity compensation, cognitive network failure prediction, dynamic planning and cross-layer optimization of software-defined networks, quality of transmission estimation, and physical layer design of optical communication systems. Recent works employing deep learning technologies are also discussed.

We experimentally demonstrate a blind scheme enabled by a frequency offset loading technique to identify the shaping factor and modulation format of probabilistically shaped 16/64/256QAM signals with various entropies. © 2020 The Author(s)

Cathepsin S (CatS), a member of cysteine cathepsin proteases, has been well studied due to its significant role in many pathological processes, including arthritis, cancer and cardiovascular diseases. CatS inhibitors have been included in D3R-GC3 for both docking pose prediction and affinity ranking, and in D3R-GC4 for binding affinity ranking. The difficulties posed by CatS inhibitors in D3R mainly come from three aspects: large size, high flexibility and similar chemical structures. We have participated in GC4; our best submitted model, which employs a similarity-based alignment docking and Vina scoring protocol, yielded Kendall’s τ of 0.23 for 459 binders in GC4. In our further explorations with machine learning, by curating a CatS specific training set, adopting a similarity-based constrained docking method as well as an arm-based fragmentation strategy which can describe large inhibitors in a locality-sensitive fashion, our best structure-based ranking protocol can achieve Kendall’s τ of 0.52 for all binders in GC4. In this exploration process, we have demonstrated the importance of training data, docking approaches and fragmentation strategies in inhibitor-ranking protocol development with machine learning.

Traditional frequency offset estimation (FOE) algorithms, such as fast Fourier transform-based FOE (FFT-FOE) and differential FOE (Diff-FOE), are no longer suitable for 32-QAM optical coherent receiver. Here, we propose a feed-forward and non-data-aided FOE algorithm for 32-QAM based on constellation rotation. After equalization, the constellation points of the second and the forth ring are selected and rotated with an angle of $\pi /4$ . Thus, those points gather along diagonal lines, which enable us to use traditional Diff-FOE for 32-QAM signal. The performance of our proposed algorithm is numerically and experimentally verified under the scenario of back-to-back transmission with dual-polarization 32-QAM signals. Our proposed FOE algorithm is more robust to amplified spontaneous emission noise in comparison with recently proposed QPSK-selection assisted FFT-FOE, while the complexities are reduced by factors of 1.7 and 3.7 in the forms of multipliers and adders, respectively.

Applications of machine learning (ML) models in optical communications and networks have been extensively investigated. For an optical wavelength-division-multiplexing (WDM) system, the quality of transmission (QoT) estimation generally depends on many parameters including the number and arrangement of WDM channels; launch power of each channel; number and distribution of fiber spans; attenuation, dispersion, and nonlinearity parameters and length of each fiber span; noise figure; gain and gain tilt of erbium-doped fiber amplifiers; transceiver noise; digital signal processing (DSP) performance; and so on. In recent years, ML-based QoT estimation schemes have gained significant attention. However, nearly all relevant works are conducted through simulations because it is difficult to obtain sufficient and high-quality datasets for training ML models. In this paper, we demonstrate completely automated generation and collection of an ultra-large-scale experimental training dataset for ML-model-based QoT estimation by automation of transceivers and optical link parameters, as well as data transfer and DSP. Implementation details and key codes of automation are presented. Artificial neural network models with one and two hidden layers are trained by the collected dataset, and brief QoT estimation results are evaluated and discussed to verify the performance and stability of the established automated system.

A motor and a generator are usually needed for converting the forms of energy between mechanical and electrical in some applications. Recently, we have proposed an energy conversion/storage device based on a unique interacting behavior between a permanent magnet and a closed superconducting coil. This device needs neither a power supply nor a motor/generator and is able to complete the conversion from mechanical energy to electrical energy or the energy storing-releasing cycle of mechanical energy → electromagnetic energy → mechanical energy. In principle, the energy conversion efficiency of this kind of device is higher than most existing energy storages. This article introduces the principle and the experimental results of a proof-of-concept prototype of the proposed device. We conducted three categories of tests on the prototype. The first was for testing the energy conversion efficiency. The second was for examining the electrical loss of the superconducting coil. The third was for exploring the delivering capacity of this kind of device. The results of these experiments demonstrate that this device is able to effectively covert mechanical energy to electrical energy with an efficiency ＞ 90%. The conversion efficiency of a mechanical energy → electrical energy → mechanical energy cycle is greater than 80%. Based on the experimental results, we believe that this kind of device is suitable for being applied in vehicle regenerative braking for energy conservation, particularly meaningful for urban rail transportations.

A linewidth-tolerant adaptive equalization scheme is proposed for M-ary offset quadrature amplitude modulation (OQAM) signal at the presence of wide laser linewidth. Initially, we present a close-form mathematical expression of OQAM signal at the presence of phase noise and derive the condition to obtain the optimal tap value of adaptive equalizer. Our theoretical investigations proves that phase noise in OQAM signal may result in variation of optimal tap value of adaptive equalizer. Consequently, conventional digital signal processing (DSP) flow that separates adaptive equalization and carrier phase recovery (CPR) into two independent modules cannot apply to OQAM signal anymore. Then, we propose a linewidth-tolerant adaptive equalization scheme that incorporate both adaptive equalizer and CPR for m-ary OQAM signal. Taking the 16-OQAM into account, we comprehensively evaluate its performance to compensate residual chromatic dispersion (CD) and polarization mode dispersion (PMD) at the presence of wide laser linewidth. Simulation shows that our proposed adaptive equalization can effectively compensate residual CD that is below 400 ps/nm without performance penalty at the presence of wide laser linewidth. In particular, a tolerance of linewidth and symbol duration products of 1 × 10 - 4 is secured under conditions of CD=400 ps/nm and DGD=10 ps, given 1-dB required-OSNR penalty at BER= 10 - 3 .

Low complexity carrier phase recovery (CPR) scheme based on vertical blind phase search (V-BPS) for M-ary offset quadrature amplitude modulation (OQAM) is proposed and numerically verified. After investigating the constellations of both even and odd samples with respect to the phase noise, we identify that the CPR can be realized by measuring the verticality of constellation with respect to different test phase angles. Then measurement without multiplication in the complex plane is found with low complexity. Furthermore, a two-stage configuration is put forward to further reduce the computational complexity (CC). Compared with our recently proposed modified blind phase search (M-BPS) algorithm, the proposed algorithm shows comparable tolerance of phase noise, but reduces the CC by a factor of 3.81 (or 3.05) in the form of multipliers (or adders), taking the CPR of 16-OQAM into account.

Structure-based drug design is critically dependent on accuracy of molecular docking scoring functions, and there is of significant interest to advance scoring functions with machine learning approaches. In this work, by judiciously expanding the training set, exploring new features related to explicit mediating water molecules as well as ligand conformation stability, and applying extreme gradient boosting (XGBoost) with Δ-Vina parametrization, we have improved robustness and applicability of machine-learning scoring functions. The new scoring function Δ(vina)XGB can not only perform consistently among the top compared to classical scoring functions for the CASF-2016 benchmark but also achieves significantly better prediction accuracy in different types of structures that mimic real docking applications.

Frequency offset (FO) is an important impairment in coherent transmission systems induced by wavelength mismatch between a free-running local oscillator (LO) and the transmit laser. A fast-Fourier-transform-based FO estimation (FFT-FOE) scheme is commonly employed. The obtained FOE value can be used to compensate FO of a relatively long duration of the signal. On the other hand, the drift of the laser frequency leads to time-varying FO, and FOE should operate in intervals to track the FO drift (FOD) for the best system performance. However, the complexity of FFT is high, and continuous operation of FFT-FOE results in huge energy consumption since it is unclear when FO drifts to an intolerable value. This Letter aims to solve such a problem: How often should FFT-FOE be operated to guarantee system performance with the lowest complexity? Here, we propose a FOD monitoring scheme based on sparse fast Fourier transform, by which optimum performance and minimum cost of FOE are achieved simultaneously. A downsampling process is further proposed to reduce the complexity of subsequent FFT-FOE. The performance is experimentally verified with 28 Gbaud coherent systems.

We propose an efficient timing/frequency synchronization assisted by sparse FFT and optimized training symbols. The proposed scheme is proved to be robust at low SNR can achieve complete frequency offset estimation range with ultra-low complexity.

We conduct a comprehensive comparative study of quality-of-transmission (QoT) estimation for wavelength-division-multiplexed systems using artificial neural network (ANN)-based machine learning (ML) models and Gaussian noise (GN) model-based analytical models. To obtain the best performance for comparison, we optimize all the system parameters for GN-based models in a brute-force manner. For ML models, we optimize the number of neurons, activation function, and number of layers. In simulation settings with perfect knowledge of system parameters and communication channels, GN-based analytical models generally outperform ANN models even though GN models are less accurate on the side channels due to the local white-noise assumption. In experimental settings, however, inaccurate knowledge of various link parameters degrades GN-based models, and ML generally estimates the QoT with better accuracy. However, ML models are temporally less stable and less generalizable to different link configurations. We also briefly study potential network capacity gains resulting from improved QoT estimators and reduced operating margins.

As one of the key indicators of signal quality in fiber communication systems, optical signal-to-noise ratio (OSNR) needs to be accurately monitored to ensure reliable network planning, operation, and reconfiguration. OSNR monitoring techniques considering only accumulated amplified spontaneous emission (ASE) noise are no longer suitable for dispersion unmanaged long-haul and dense wavelength division multiplexing (WDM) systems, where the contribution of fiber nonlinearity to total SNR cannot be neglected. In this paper, we propose a modulation-format-transparent, accurate joint linear and nonlinear noise monitoring scheme based on calculation of correlation between two spectral components at the upper and lower sideband of the signal spectrum. Different characteristics of flat linear noise spectrum and non-flat nonlinear noise spectrum are used to distinguish the influences on the correlation value from both noise sources. Simulation results show that the proposed scheme can accurately monitor SNRlinear and SNRnonlinear within a wide launch power range from −5 dBm to 5 dBm per channel for multi-channel WDM systems with a 915-km single mode fiber (SMF) link. The performance of the proposed scheme is further experimentally verified in up-to-7 channel WDM systems over a 915 km SMF link.

We propose a blind and fast modulation format identification (MFI) enabled by the digital frequency-offset (FO) loading technique for hitless coherent transceiver. Since modulation format information is encoded to the FO distribution during digital signal processing (DSP) at the transmitter side (Tx), we can use the fast Fourier transformation based FO estimation (FFT-FOE) method to obtain the FO distribution of individual data block after constant modulus algorithm (CMA) pre-equalization at the receiver side, in order to realize non-data-aided (NDA) and fast MFI. The obtained FO can be also used for subsequent FO compensation (FOC), without additional complexity. We numerically investigate and experimentally verify the proposed MFI with high accuracy and fast format switching among 28 Gbaud dual-polarization (DP)-4/8/16/64QAM, time domain hybrid-4/16QAM, and set partitioning (SP)-128QAM. In particular, the proposed MFI brings no performance degradation, in term of tolerance of amplified spontaneous emission (ASE) noise, laser linewidth, and fiber nonlinearity. Finally, a hitless coherent transceiver enabled by the proposed MFI with switching-block of only 2048 symbols is demonstrated over 1500 km standard single mode fiber (SSMF) transmission.

We propose a blind and fast modulation format identification enabled by frequency-offset loading technique for hitless flexible transceiver. Arbitrary format switching among 28-Gbaud DP-4/8/16/64QAM, hybrid-4/16QAM, and SP-128QAM can be accurately identified with only 2048 symbols.

Due to the non-rectangular distribution of the constellation points, traditional fast Fourier transform based frequency offset estimation (FFT-FOE) is no longer suitable for 32-QAM signal. Here, we report a modified FFT-FOE technique by selecting and digitally amplifying the inner QPSK ring of 32-QAM after the adaptive equalization, which is defined as QPSK-selection assisted FFT-FOE. Simulation results show that no FOE error occurs with a FFT size of only 512 symbols, when the signal-to-noise ratio (SNR) is above 17.5 dB using our proposed FOE technique. However, the error probability of traditional FFT-FOE scheme for 32-QAM is always intolerant. Finally, our proposed FOE scheme functions well for 10 Gbaud dual polarization (DP)-32-QAM signal to reach 20% forward error correction (FEC) threshold of BER=2×10sup-2/sup, under the scenario of back-to-back (B2B) transmission.

We experimentally demonstrate a novel carrier phase estimation (CPE) scheme based on real distance-blind phase search (RD-BPS) for M-QAM. Compared with traditional BPS, the RD-BPS is verified to reduce the complexity without performance degradation.

For hitless flexible coherent transceivers based next-generation agile optical network, efficient modulation format identification (MFI) is an essential element in digital signal processing (DSP) flow at the receiver-side (Rx). In this paper, we propose a blind and fast MFI scheme with high identification accuracy at low optical signal-to-noise ratio (OSNR) regime. This is achieved by first raising the signal to the 4th power and calculate the peak-to-average power ratio (PAPR) of the corresponding spectra to distinguish 32 quadrature amplitude modulation (QAM) from quadrature phase shift keying (QPSK), 16 and 64QAM signals. Then, followed by iterative partition schemes to remove signals with phase ±π4,±3π4 (or QPSK-like phases) based on the signal amplitudes, the PAPR of the remaining signals is calculated to distinguish the other three formats. Additionally, by frequency offset (FO) pre-compensation, the spectrum can be obtained using sparse-fast-Fourier-transform (S-FFT), which greatly reduces the total complexity. The MFI performance is numerically and experimentally investigated by 28 Gbaud dual-polarization (DP) coherent optical back-to-back (B2B) and up to 1500 km standard single mode fiber (SSMF) transmission system using QPSK, 16QAM, 32QAM, and 64QAM. Results show that high identification accuracy can be achieved, even when OSNR is lower than that required for the 20% forward error correction (FEC) threshold of BER=2×10-2 for each format. Furthermore, fast format switching between 64QAM-32QAM and 32QAM-16QAM are demonstrated experimentally for B2B scenario and 900 km SSMF with the proposed MFI technique, respectively.

We propose and verify a carrier phase estimation algorithm for M-ary quadrature amplitude modulation (M-QAM) based on blind phase search (BPS). By replacing the complicated squared distance calculations with a simplified measurement in the real plane, the inherent computation complexity (CC) is reduced without performance degradation.

We present dual-polarization complex-weighted, decision-aided, maximum-likelihood algorithm with superscalar parallelization (SSP-DP-CW-DA-ML) for joint carrier phase and frequency-offset estimation (FOE) in coherent optical receivers. By pre-compensation of the phase offset between signals in dual polarizations, the performance can be substantially improved. Meanwhile, with the help of modified SSP-based parallel implementation, the acquisition time of FO and the required number of training symbols are reduced by transferring the complex weights of the filters between adjacent buffers, where differential coding/decoding is not required. Simulation results show that the laser linewidth tolerance of our proposed algorithm is comparable to traditional blind phase search (BPS), while a complete FOE range of ± symbol rate/2 can be achieved. Finally, performance of our proposed algorithm is experimentally verified under the scenario of back-to-back (B2B) transmission using 10 Gbaud DP-16/32-QAM formats.

We experimentally compare QoT estimations for WDM systems using Machine Learning(ML) and GN-based analytical models. ML estimates the side channels with better accuracy but is temporally less stable and less generalizable to different link configurations.