Informative experimental data are critical for developing predictive combustion kinetic models. Bayesian Sequential Experimental Design (BSED) provides a principled framework to identify experimental conditions that maximize the Expected Information Gain (EIG). However, its application to kinetic model optimization is often limited by high computational cost and potential model misspecification which leads to suboptimal designs. To address these challenges, this study introduces the Surrogate Accelerated BSED (SABSED) framework and demonstrates its implementation on a semi-automated Jet-Stirred Reactor (JSR) platform. SABSED is a surrogate-assisted nonlinear design framework that avoids input-output linearization approximations and improves experimental efficiency through enhanced information gain. The framework integrates multiple strategies to enhance both efficiency and robustness. For efficiency, it employs Artificial Neural Networks (ANNs) trained on multi-scenario datasets to accelerate simulations, utilizes reverse Kullback-Leibler divergence for rapid EIG evaluation, and applies ANN-based Hamiltonian Monte Carlo for efficient Bayesian inference. For robustness, it introduces a heteroscedastic Gaussian Process Regression (GPR) surrogate model to define a modified EIG criterion (EIG_GPR) that accounts for prediction-measurement difference, improving real-world design reliability. Validation using ammonia combustion data, including ignition delay times (IDT), laminar burning velocities (LBV), and species profiles from JSR experiments, demonstrates that both EIG and EIG_GPR substantially reduce predictive errors and uncertainties within 20 iterations. In the LBV case, EIG_GPR achieved the target error threshold with five times fewer iterations than EIG. For newly designed JSR experiments, EIG_GPR accelerated the design and optimization process by 15 times, completing the design within 30 s on a single-core CPU. Overall, the SABSED framework significantly enhances the efficiency of developing combustion reaction kinetic models, paving the way for autonomous experimentation and automatic model optimization.

While kinetic mechanisms play a pivotal role in simulating complex combustion problems, their extended scale often results in prohibitive computational cost, particularly when integrated with computational fluid dynamics simulations. This paper introduces the Directed Relation Graph Species Rank (DRGSR) algorithm, an efficient mechanism reduction technique designed to retain the essential species and reaction pathways while minimizing computational demands. Specifically, it incorporates a two-step approach: the first utilizes a directed relation graph to map species interactions and transform the kinetic information into a graph structure, and the second employs the PageRank algorithm and directed interaction coefficients to rank species based on their importance within the network. The DRGSR algorithm is validated through case studies involving both large- and small-scale, high-temperature and low-temperature mechanisms, specifically focusing on the ignition delay times for ethylene (C2H4) and n-heptane (C7H16). The algorithm demonstrates superior performance in reducing the number of species significantly while maintaining accuracy; for ethylene, it retains only 31 species with an error under 8 %, while for n-heptane, it achieves comparable precision with fewer species compared to existing methods. The validation is extended to predicting the laminar flame speeds, and further affirms the algorithm's reliability and generalizability. A comparative analysis of the computational cost reveals that the DRGSR algorithm not only is less time-consuming, but it also simplifies the reduction process by eliminating the iterative threshold adjustments required by methods such as Directed Relation Graph (DRG), Directed Relation Graph with Error Propagation (DRGEP) and Directed Relation Graph with Error Propagation and Sensitivity Analysis (DRGEPSA). These findings indicate that the DRGSR algorithm offers a robust, efficient and reliable approach for kinetic mechanism reduction, suitable for wide ranges of engineering applications.

In response to the interest in nitrogen-containing compounds as energetic materials, an experimental and kinetic study on the low-temperature oxidation of three butyl nitrites isomers, namely n-butyl (NBN), isobutyl (IBN), and tert-butyl (TBN) was performed. By measuring their ignition delays in a rapid compression machine (RCM) under 5-15 bar at temperatures from 550 to 630 K, a two-stage ignition behavior was observed for all the three nitrites, with the first-stage delays of TBN being shorter than those of NBN and IBN. A detailed kinetic mechanism was constructed and validated against the experimental data, and the production rate was analyzed to explain the first-stage ignition behavior. Specifically, the N-O bond dissociation reaction initiated the consumption of butyl nitrites isomers in all cases studied, which produced NO and different butoxy radicals (C4H9O). In the case of TBN, the decomposition of TC4H9O produces CH3 in the first-stage ignition. The abundant CH3 radical reacts with NO2 to produce CH3O, which further yields HO2 and CH2O through the reaction with O2. The inert HO2 radical is converted to OH through the reaction HO2 + NO = OH + NO2, resulting in the first-stage ignition. Meanwhile, the decomposition of PC4H9O and IC4H9O produces n-propyl and i-propyl radicals, respectively, in the cases of NBN and IBN. The reaction sequences of n-propyl and i-propyl radicals produce less HO2 radicals compared with that in TBN, leading to longer first-stage ignition time.

The construction of surrogate models is an essential step in the uncertainty quantification of combustion reaction kinetics. These models create a mapping between inputs and outputs of combustion kinetics simulations, thereby replacing the time-consuming numerical simulations of reaction kinetics and significantly lowering the computational costs for uncertainty quantification. However, in applications such as experimental design that require repeated construction of surrogate models under multiple operating conditions, the associated computational burden becomes substantial and can even limit the feasibility of the entire task. It is therefore essential to investigate cost-efficient surrogate model construction methods. Drawing inspiration from image classification in computer vision, this work introduces a meta-learning-assisted approach to efficiently construct surrogate models by leveraging the intrinsic shared features among them. By learning from a limited set of training tasks, the approach facilitates rapid creating surrogate models for new conditions with fewer samples. This is particularly beneficial for reducing computational costs since the most significant expense comes from the generation of original samples. The method has been tested in ammonia-hydrogen combustion targeting ignition delay time and laminar burning velocity. Results show that the efficiency of the surrogate model construction can be improved by a factor of eight for individual new conditions, and the total computational costs across the entire condition range can be reduced to 29 % and 37 % of the original values for the two prediction targets, respectively. Notably, dual pretraining across both prediction targets further enhances model performance. The meta-learning-assisted surrogate model construction approach is applicable across a broad range of operating conditions, requiring only minimal additional pretraining costs while offering flexible precision control based on task-specific requirements.

Ammonia has emerged as a highly promising zero-carbon energy carrier in recent years. To address its combustion challenges, blending with hydrogen has been proposed as an effective solution. This approach necessitates a comprehensive understanding of the underlying reaction kinetics for practical implementation. The extensive experimental data available on NH3/H2 combustion offers the potential for advancing chemical kinetics understanding through model optimization. However, effectively utilizing the extensive dataset poses significant challenges, primarily due to prohibitively high computational costs when processing large volumes of data for model optimization. Moreover, indiscriminate use of all available data without proper quality assessment is inadvisable, given the potential issues of data inconsistency. To comprehensively extract kinetic information from existing experimental dataset and facilitate efficient model development, this study implements a global-sensitivity-based clustering approach. The 2786 NH3/H2 experimental data are categorized into 40 distinct clusters. Representative exemplars from each cluster are selected to construct a streamlined yet information-dense dataset, which is subsequently employed for model optimization, resulting in a substantial enhancement of the model's predictive performance. Although information redundancy exists in current datasets, certain aspects of reaction kinetics may remain unexplored due to limitations in existing experimental data. Within the range of experimental conditions that current experimental equipment can cover, there are still many unexplored condition regions. To investigate the potential impact of conducting experiments under these new conditions on advancing NH3/H2 combustion modeling, we systematically extend the dataset to incorporate 6695 theoretically feasible experimental conditions. Notably, eight previously insensitive reactions demonstrate significant sensitivity within this expanded framework. Furthermore, eleven additional prospective experimental conditions exhibiting high sensitivity to these eight parameters are identified and incorporated into the existing informative dataset. Subsequent model optimization utilizing this enhanced dataset of 51 elements yields significantly improved performance. This study demonstrates that carefully-designed future experiments can provide novel insights into the reaction kinetics of ammonia-hydrogen combustion kinetics, as well as offering valuable guidance for future experimental investigations.

This study presents a novel method for optimizing combustion kinetic mechanisms using variational inference (VI), addressing the longstanding challenge of balancing accuracy and computational efficiency. The primary focus is on the uncertainty quantification and minimization of pre-exponential factors in combustion reactions, a critical aspect of kinetic mechanism optimization. The initial phase involves identifying the most sensitive parameters through local sensitivity analysis, followed by the application of artificial neural networks (ANN) as surrogate models for each experimental condition. This approach significantly accelerates the optimization process while maintaining gradient propagation essential for variational inference. A key advancement in this research is the implementation of variational inference in place of traditional Markov chain Monte Carlo (MCMC) methods. This shift not only results in a marked improvement in the computational speed, but it also maintains a high degree of accuracy in parameter optimization. This is illustrated through a detailed comparison of posterior distributions and uncertainty constraints obtained via both MCMC and variational inference methods, employing a methanol mechanism as a test case. Moreover, the inclusion of covariance in the variational inference effectively addresses the over-constraint of parameter uncertainties observed in earlier methods, offering a more realistic depiction of parameter interdependencies. The results show complete consistency in covariance matrices inferred from both MCMC and variational inference, validating the accuracy of the new probabilistic model. In terms of computational efficiency, the variational inference method demonstrates an enhancement of more than three orders of magnitude compared to the ANN-MCMC method. This significant reduction in time cost paves the way for applying the method to larger-scale mechanisms and data sets.

Unsupervised cross-modal hashing (UCMH) has been commonly explored to support large-scale cross-modal retrieval of unlabeled data. Despite promising progress, most existing approaches are developed on convolutional neural network and multilayer perceptron architectures, sacrificing the quality of hash codes due to limited capacity for excavating multi-modal semantics. To pursue better content understanding, we break this convention for UCMH and delve into a transformer-based paradigm. Unlike naïve adaptations via backbone substitution that overlook the heterogeneous semantics from transformers, we propose a multi-granularity learning framework called hugging to bridge the modality gap. Specifically, we first construct a fine-grained semantic space composed of a series of aggregated local embeddings that capture implicit attribute-level semantics. In the hash learning stage, we innovatively incorporate fine-grained alignment with these local embeddings to enhance global hash code alignment. Notably, this fine-grained alignment only facilitates robust cross-modal learning without complicating global hash code generation at test time, thus fully maintaining the high efficiency of hash-based retrieval. To make the most of fine-grained information, we further propose a differentiable optimized quantization algorithm and extend our framework to hugging ^+ . This variant neatly integrates quantization learning into the fine-grained alignment during training, producing quantization codes of local embeddings as a gift at test time, which can augment the retrieval performance through an efficient reranking stage. We instantiate simple baselines with contrastive learning objectives for hugging and hugging ^+ , namely HuggingHash and HuggingHash ^+ . Extensive experiments on 4 text-image retrieval and 2 text-video retrieval benchmark datasets show the competitive performance of HuggingHash and HuggingHash ^+ against state-of-the-art baselines. More encouragingly, we also validate that hugging and hugging ^+ are flexible and effective across various baselines, suggesting their universal applicability in the realm of UCMH.

In the realm of combustion kinetic modeling, the norm involves employing thousands of reactions to delineate the chemical conversion of hundreds of species. Notably, theoretically predicted rate coefficients and branching ratios, derived through the RRKM/master equation (ME) model, play an increasing role in kinetic modeling. Thus minimizing the uncertainty of theoretical prediction across wide working conditions is crucial to refine a kinetic model. The present study takes ethyl (C2H5) + oxygen (O2) reaction system to show that combined forward and reverse uncertainty analysis can be used to further constrain calculated rate coefficients and branching ratios, which were already calculated by high-level quantum chemistry methods. Forward global uncertainty analysis with the artificial neural network-high dimensional model representation (ANN-HDMR) method is employed to select key parameters affecting total rate coefficients of C2H5 + O2 and branching ratios of C2H5 + O2 = C2H4 + HO2 (C1). Reverse uncertainty analysis with Bayesian method was then applied to refine the key input parameters based on experimental data at working conditions selected by sensitivity entropy. Although the target RRKM/ME model system was built on high level theoretical calculations, the combined forward and reverse uncertainty analyses are still able to reduce uncertainties of predicted total rate coefficients of C2H5 + O2 and branching ratios for C1 across a wide range of working conditions. Specifically, the uncertainties of total rate coefficient and C1 branching ratio have been reduced from 1.46 and 1.52 to 1.30 and 1.36 at 298 K and 1 Torr. The analysis process proposed in the present work effectively extrapolates the constraint ability of accurate measured data at one condition to wide working conditions based on the RRKM/ME model.

Bayesian inference plays a pivotal role in optimizing reaction kinetics models, with Markov Chain Monte Carlo (MCMC) methods being a critical implementation. However, MCMC is constrained with low acceptance rates in high-dimensional scenarios. Conversely, Hamiltonian Monte Carlo (HMC) has high acceptance rates, albeit at the expense of computing first-order log-posterior gradients, thus limiting its application in model optimization. This study introduces an Artificial Neural Network-based HMC (ANNHMC) approach, wherein gradient computation is accelerated through automatic differentiation. Unlike previous models that depend on early HMC samples for training surrogate models to predict the Hamiltonian, our ANNHMC employs prediction samples for training an ANN. This ANN predicts targets used in the Hamiltonian computation, hence significantly streamlines the training process. The experimental data of ignition delay time and laminar flame velocity are utilized to optimize the methanol, ammonia and JetSurF1.0 models. The optimization results show that ANNHMC outperforms ANN-MCMC, enhancing the sampling efficiency by three to four orders of magnitude after neural network training. Additionally, we investigate the influence of active parameter selection on model optimization under uncertainty constraints. The global sensitivity analysis typically identifies fewer total parameters than the local sensitivity analysis, resulting in more stringent parameter constraints. This can amplify the sampling challenge for ANNHMC as some parameters approach boundary limits, thereby increasing the overall sampling time. In addition, more stringent parameter constraints may potentially limit the model performance on the optimization dataset. Appropriately increasing the number of active parameters can improve the model prediction accuracy, while maintaining prediction uncertainty within the limits of experimental uncertainty. The ANNHMC demonstrates its unique advantages in reaction kinetics model optimization.

In the field of video captioning, recent works usually focus on multi-modal video content understanding, in which transcripts are extracted from speech and are often adopted as an informational supplement. However, most existing works only consider transcripts as a supplementary modality, neglecting their potential in capturing high-level semantics, such as multi-modal topics. In fact, transcripts, as a textual attribute derived from the video, reflect the same high-level topics as the video content. Nonetheless, how to resolve the heterogeneity of multi-modal topics is still under-investigated and worth exploring. In this paper, we introduce a contrastive topic-enhanced network to consistently model heterogeneous topics, that is, inject an alignment module in advance, to learn a comprehensive latent topic space and guide caption generation. Specifically, our method includes a local semantic alignment module and a global topic fusion module. In the local semantic alignment module, a fine-grained semantic alignment at the clip-sentence granularity reduces the semantic gap between modalities. Extensive experiments have verified the effectiveness of our solution.

Ammonia (NH3) has shown promise as a carbon-neutral fuel, and blending NH3 with methane (CH4) can improve its combustion characteristics. This study investigated the auto-ignition characteristics and kinetic modeling of NH3/CH4 mixtures using a rapid compression machine (RCM). Ignition delay times were measured under various conditions, including three different NH3/CH4 ratios, pressures of 15 and 25 bar, temperatures ranging from 910 to 1272 K, and equivalence ratios of 0.5 and 1.0. The results showed the addition of CH4 significantly promoted NH3 ignition. An intriguing phenomenon of pre-ignition pressure rise was noted in blends containing 25 % CH4, which cannot be replicated by previous models. Time-resolved species measurements were conducted using a fast sampling system coupled with gas chromatography. The measurements revealed distinct early consumption of CH4 compared to NH3 in blends exhibiting pre-ignition pressure rise. A kinetic model was developed for NH3/CH4 combustion with emphasis on key "CN" reaction pathways constrained by experimental results. Sensitivity and reaction path analyses highlighted the significance of these "CN" reactions in determining the varying ignition characteristics. Additional simulations systematically assessed factors influencing auto-ignition behavior in NH3/CH4 mixtures. Notably, CH4 concentration emerged as the primary parameter affecting the ignition behavior. Lower CH4 concentrations led to NH3 chemistry primarily controlling the overall ignition process, which could cause early CH4 oxidation and potential gradual pre-ignition pressure rise. Conversely, higher CH4 content facilitated its oxidation to trigger the oxidation of NH3, thereby narrowing the time gap between the initiation of consumption for the two fuels.

To minimize the uncertainty of the parameters in combustion kinetics models, Bayesian methods are commonly used for uncertainty constraints based on experimental data. With the rapid and substantial growth of experimental data, using all the experimental data for optimization is not only redundant and time-consuming, but it could also lead to data consistency problems. In this work, the global sensitivitybased affinity propagation method (GSAP) is proposed to cluster experimental datasets and to select representative experimental conditions. Specifically, the global sensitivity coefficient is first obtained through an analysis to characterize the sources of uncertainty in the kinetic model under different experimental conditions. The similarity coefficient, which is defined based on the global sensitivity, measures the resemblance between two experimental conditions. By exchanging messages calculated from similarity, affinity propagation enables the experimental dataset to be automatically clustered into several classes without specifying the number of classes in advance. This method innovatively introduces the consideration of model and experimental uncertainty under different conditions to obtain better optimization results. The correctness and effectiveness of the method are validated through clustering and optimizing on a laminar flame speed dataset of common C 0 -C 4 fuels. The dataset consisting of 288 experimental conditions has been automatically clustered into 27 categories, and an exemplar of each category is given. These exemplary conditions reflect the dominant chemistry behind their cluster. At the same time, these conditions have larger model prediction uncertainty and smaller experimental uncertainty to provide better Bayesian constraints. The uncertainty of the model parameters after Bayesian optimization is effectively constrained. The average uncertainty of model predictions across the dataset is reduced from 30 % to 10 % using only 27 exemplar conditions for optimization. While selecting experimental data for model optimization, the clustering strategies provided by this method also, in turn, help understand its underlying chemical essence.(c) 2023 The Combustion Institute. Published by Elsevier Inc. All rights reserved.

This study investigated the two-stage ignition behavior of NH 3 /H 2 mixtures using a rapid compression machine (RCM). The ignition delay times and first-stage ignition delay times of NH 3 /H 2 blends were measured at temperatures ranging from 881 to 1127 K, pressures of 15 and 25 bar, and equivalence ratios ( & phi;) of 1.0 and 1.5. Experimental results showed that H 2 had a significant promotional effect on the ignition of NH 3 , and two-stage ignition behavior was observed in some test mixtures. Time-resolved species concentrations were recorded using the gas chromatography (GC) method during the single- and two-stage ignition process. Species evolution suggested that the consumption of NH 3 and H 2 separated during the two-stage ignition process. The oxidation of H 2 primarily occurred at the first-stage ignition point, while NH 3 oxidation occurred at the total ignition point. A kinetic model was developed to predict the twostage ignition behavior of NH 3 /H 2 and the species profiles. Model analysis showed that H 2 played a key role in initiating the oxidation process and contributed to the early heat release. When H 2 content was high (e.g. 50%), its oxidation led to the simultaneous total oxidation of NH 3 , resulting in single-stage ignition characteristics. However, when a small amount of H 2 was present (e.g. 10%), its oxidation only partially consumed NH 3 , leading to the two-stage ignition behavior. Sensitivity analyses indicated that the co-oxidation of fuels during the first-stage ignition was primarily dominated by H 2 oxidation chemistry, while NH 3 oxidation became dominant during the following total ignition stage as H 2 was completely consumed. Additional model simulations revealed that the ignition behavior of NH 3 /H 2 mixtures is strongly influenced by temperature and pressure. A distinct threshold for temperature and pressure was identified, demarcating the transition between single-stage and two-stage ignition phenomena. Moreover, the fraction of H 2 in the mixture had a significant impact on the ignition behavior, with higher fractions leading to increased intensity of the first-stage ignition and closer proximity to the total ignition point. However, when the H 2 fraction exceeds a certain threshold, two-stage ignition behavior disappears.& COPY; 2023 The Combustion Institute. Published by Elsevier Inc. All rights reserved.

To extend the temperature for propane ignition to a lower region (< 680 K), ozone (O 3 ) was used as an ignition promoter to investigate the low-temperature chemistry of propane. Ignition delay times for propane containing varying concentrations of O 3 (0, 100, and 1000 ppm) were measured at 25 bar, 654-882 K, and equivalence ratios of 0.5 and 1.0 in a rapid compression machine (RCM). Species profiles during propane ignition with varying O 3 concentrations were recorded using a fast sampling system combined with a gas chromatograph (GC). A kinetic model for propane ignition with O 3 was developed. O 3 shortened ignition delay times of propane significantly, and the NTC behavior was weakened. O atoms released from O 3 re-acted with propane through hydrogen abstraction reactions, which led to the fast production of OH radi-cals. The following oxidation of fuel radicals generated additional OH radicals. Consequently, the inhibition caused by the slow chemistry of hydrogen peroxide (H2O2) in the NTC region was weakened in the presence of O 3 . Experimental results with O 3 addition can provide extra constraints on the low-temperature chem-istry of propane. Species profiles during propane ignition at 730 K with 1000 ppm O 3 addition showed the production of propanal (C2H5CHO), acetone (CH3COCH3), and acetaldehyde (CH3CHO) was promoted significantly. Model analyses indicated that O 3 shifted the oxidation temperature of propane to a lower re-gion, in which reactions of ROO radicals (NC3H7O2 and IC3H7O2) tend to generate RO radicals (NC3H7O and IC3H7O). The promotion of RO radicals led to the fast production of C2H5CHO, CH3COCH3, and CH3CHO. The corresponding species profile highlighted the reaction relevant to ROO and RO radicals (NC3H7O + O 2 = C2H5CHO + HO2 and 2 IC3H7O2 = 2 IC3H7O + O 2 ). Rate constants of these reac-tions were updated, which can potentially improve the performance of the core mechanism under lower temperatures and provide references for model development of larger hydrocarbons. & COPY; 2022 The Combustion Institute. Published by Elsevier Inc. All rights reserved.

Experimental data is essential for the improvement of combustion kinetic models. Experimental design based on model analysis results can screen optimal experimental conditions with maximum information content. However, the computational cost of designing experiments by enumeration becomes unaffordable when an enormity of conditions with different temperatures/pressures/mixtures are to be investigated. An approach to facilitate the efficient discovery of optimal experimental conditions based on the genetic algorithm (GA) is proposed in this work. This approach regards the task of experimental design as an optimization problem to minimize an objective function that measures the information content provided by an experiment. The sensitivity entropy and surrogate model similarity are combined to form the objective function of optimization. Three designs of dimethyl ether experiments are provided to demonstrate the approach. The first case utilizes a benchmark for optimal experiments to validate the effectiveness of GA. The results show that GA can achieve better design results than the traditional enumeration strategy with less than 10% computational cost. The second case illustrates how GA is applied in the design of multiple experiments. The last one is an application in designing multiple experiments of various types, including ignition, species measurements in a jet-stirred reactor (JSR) and a plug flow reactor (PFR). The model parameters are calibrated with the designed experimental data using a Bayesian-based optimization approach. The uncertainties of model parameters are significantly reduced after the optimization.

Surrogate models are often used to accelerate the uncertainty quantification (UQ) of chemical kinetic models. However, the construction of surrogate models usually requires the repetitive generation of high-fidelity input-output samples, which is time-consuming. In this work, we utilize a multi-fidelity neural network to speed up the surrogate model construction, yielding a multi-fidelity neural network-based surrogate model (MFNNSM). MFNNSM consists of two separate neural networks. The first neural network learns from high and low-fidelity samples, and then generates samples to train the second neural network for the subsequent UQ analysis. Based on the fact that the uncertainty similarity (sensitivity-based similarity) does exist between the predictions from the reduced and detailed combustion kinetics models, or between the predictions under different conditions, MFNNSM can use reduced model predictions as low-fidelity samples to generate detailed model predictions as the high-fidelity samples, or can transfer samples across different simulations conditions. To demonstrate the MFNNSM method, the ignition delay time (IDT) and laminar flame velocity (LFV) of methanol and n-decane are chosen as model prediction targets for UQ analysis. The results show that MFNNSM can achieve acceleration factors up to 6, by utilizing reduced model samples to generate detailed model samples under the same conditions, and when reusing the reduced model samples under different conditions, the acceleration factor can increase to 10. In addition, the influences of the relative error of the reduced models and the uncertainty similarity coefficients between combustion conditions on the performance of MFNNSM are further discussed, providing the guidance for future applications of MFNNSM.

Deep insights into the combustion kinetics of ammonia (NH3) can facilitate its application as a promis-ing carbon-free fuel. Due to the low reactivity of NH3, experimental data of NH3 combustion can only be obtained within a limited range. In this work, nitrous oxide (N2O) and hydrogen (H 2 ) were used as addi-tives to investigate NH3 auto-ignition in a rapid compression machine (RCM). Ignition delay times for NH3, NH3/N2O blends, and NH3/H2 blends were measured at 30 bar, temperatures from 950 to 1437 K. The addi-tion of N2O and H 2 ranged from 0 to 50% and 0 to 25% of NH3 mole fraction, respectively. Time-resolved species profiles were recorded during the auto-ignition process using a fast sampling system combined with a gas chromatograph (GC). An NH3 combustion model was developed, in which the rate constants of key reactions were constrained by current experimental data. The addition of N2O affected the ignition of NH3 primarily through the decomposition of N2O (N2O ( + M) = N 2 + O ( + M), R1) and direct reaction between N2O and NH2 (N2H2 + NO = NH2 + N2O, R2). The rate constant of R2 was constrained effectively by experimental data of NH3/N2O mixtures. Two-stage ignition behaviors were observed for NH3/H2 mixtures, and the corresponding first-stage ignition delay times were reported for the first time. Experimental species profiles suggested the first-stage ignition resulted from the consumption of H 2 . The oxidation of H 2 pro-vided extra HO2 radicals, which promoted the production of OH radicals and initiated first-stage ignition. Reactions between HO2 radicals and NH3/NH2 dominated the first-ignition delay times of NH3 /H 2 mixtures. Moreover, the first-stage ignition led to the fast production of NO2, which acted as a key intermediate and affected the following total ignition. Consequently, the reaction NH2 + NO2 = H2NO + NO (R3) was constrained by total ignition delay times.& COPY; 2022 The Combustion Institute. Published by Elsevier Inc. All rights reserved.

Recognizing that the calibration of octane number (ON) of a fuel by standard experimental testings is often challenging due to the lack of samples and the complexity of the experimental operating conditions, we propose herein the use of convolutional neural network (CNN) method for its prediction based on the time-resolved information contained in the profiles of some small combustion species (e.g., OH, HO2, CH2O) involved in constant volume autoignition. The approach first pre-processes the species profiles obtained from experiments or simulations as input parameters and then uses convolutional neural networks for feature extraction. The obtained features are concatenated with the corresponding temperature, pressure, and ignition delay time and fed into a multilayer perceptron neural network for ON prediction. The method is validated on data sets consisting of fuel blends and various single components, including alkanes, esters, alcohols, etc. Results show that the method exhibits a high accuracy for predicting the ON of not only single component fuels but also fuel mixtures with a mean absolute error of less than 2, and that parameter sharing allows the neural network to use few parameters while extracting some high-level semantic features. Furthermore, since the input information is some common small species, the method can make predictions for almost any fuel, especially for fuel blends whose information on physical parameters and molecular structure is not available.

In addition to relevance, diversity is an important yet less studied performance metric of cross-modal image retrieval systems, which is critical to user experience. Existing solutions for diversity-aware image retrieval either explicitly post-process the raw retrieval results from standard retrieval systems or try to learn multi-vector representations of images to represent their diverse semantics. However, neither of them is good enough to balance relevance and diversity. On the one hand, standard retrieval systems are usually biased to common semantics and seldom exploit diversity-aware regularization in training, which makes it difficult to promote diversity by post-processing. On the other hand, multi-vector representation methods are not guaranteed to learn robust multiple projections. As a result, irrelevant images and images of rare or unique semantics may be projected inappropriately, which degrades the relevance and diversity of the results generated by some typical algorithms like top-k. To cope with these problems, this paper presents a new method called CoLT that tries to generate much more representative and robust representations for accurately classifying images. Specifically, CoLT first extracts semantics-aware image features by enhancing the preliminary representations of an existing one-to-one cross-modal system with semantics-aware contrastive learning. Then, a transformer-based token classifier is developed to subsume all the features into their corresponding categories. Finally, a post-processing algorithm is designed to retrieve images from each category to form the final retrieval result. Extensive experiments on two real-world datasets Div400 and Div150Cred show that CoLT can effectively boost diversity, and outperforms the existing methods as a whole (with a higher F1 score).

The task of cross-modal image retrieval has recently attracted considerable research attention. In real-world scenarios, keyword-based queries issued by users are usually short and have broad semantics. Therefore, semantic diversity is as important as retrieval accuracy in such user-oriented services, which improves user experience. However, most typical cross-modal image retrieval methods based on single point query embedding inevitably result in low semantic diversity, while existing diverse retrieval approaches frequently lead to low accuracy due to a lack of cross-modal understanding. To address this challenge, we introduce an end-to-end solution termed variational multiple instance graph (VMIG), in which a continuous semantic space is learned to capture diverse query semantics, and the retrieval task is formulated as a multiple instance learning problems to connect diverse features across modalities. Specifically, a query-guided variational autoencoder is employed to model the continuous semantic space instead of learning a single-point embedding. Afterward, multiple instances of the image and query are obtained by sampling in the continuous semantic space and applying multihead attention, respectively. Thereafter, an instance graph is constructed to remove noisy instances and align cross-modal semantics. Finally, heterogeneous modalities are robustly fused under multiple losses. Extensive experiments on two real-world datasets have well verified the effectiveness of our proposed solution in both retrieval accuracy and semantic diversity.