Soundar Kumara

Small-scale, multispecific fisheries in the Gulf of California face significant challenges including limited species-specific catch data, uncertainty about climate change impacts, and insufficient biological information needed for traditional deterministic models. These knowledge gaps hamper efforts to forecast future conditions and develop appropriate management strategies accurately. The complexity of these multi-species fisheries, combined with data scarcity for many target species, creates substantial barriers to quantifying and addressing climate vulnerability. Deep learning approaches offer a promising alternative by leveraging available data to identify patterns and project trends despite these limitations, providing valuable insights for fisheries management in data-poor contexts. Here, we apply a Mixture of Expert, a deep learning forecasting models for small-scale, multi-specific fisheries in the Gulf of California under future climate change scenarios. Results show varied responses across marine habitats, with reef and benthic fish projected to experience substantial declines (-12.46% and -9.37%) during the 2050s-2060s, followed by recovery in the 2070s-2080s. Economic implications are significant, with reef fish facing projected losses of $1.2 million by the 2050s before recovering by the 2080s. Shapley Additive Explanations (SHAP) analysis was applied to evaluate the importance of features for each predictive model, the analysis revealed the effects of the temperature in different depths for each fishery, and the sensitive analysis pointed to the magnitude of the effect. Our findings suggest that climate impacts will not be uniform across the Gulf, necessitating region-specific management approaches and highlighting the value of maintaining diverse fishing portfolios to enhance resilience against climate-driven changes.

The amount of data collected for sensing tasks in scientific computing is based on the Shannon-Nyquist sampling theorem proposed in the 1940s. Sensor data generation will surpass 73 trillion GB by 2025 as we increase the high-fidelity digitization of the physical world. Skyrocketing data infrastructure costs and time to maintain and compute on all this data are increasingly common. To address this, we introduce a selective learning approach, where the amount of data collected is problem dependent. We develop novel shift-invariant and spectrally stable neural networks to solve real-time sensing problems formulated as classification or regression problems. We demonstrate that (i) less data can be collected while preserving information, and (ii) test accuracy improves with data augmentation (size of training data), rather than by collecting more than a certain fraction of raw data, unlike information theoretic approaches. While sampling at Nyquist rates, every data point does not have to be resolved at Nyquist and the network learns the amount of data to be collected. This has significant implications (orders of magnitude reduction) on the amount of data collected, computation, power, time, bandwidth, and latency required for several embedded applications ranging from low earth orbit economy to unmanned underwater vehicles.

Metal Additive Manufacturing (MAM) produces complex, part geometries from a variety of materials in powder and wire form. Due to complexities of MAM processes that create those geometries, especially powder bed fusion, quality assurance, and qualification remain an ongoing challenge. Quality assurance involves assessing the quality of a part’s geometry, surface finish, and mechanical properties. Currently, quality assurance is not easily achieved due to variations in the powder inputs, the MAM process itself, and environmental factors, such as temperature. Many efforts are underway to develop a new quality system that includes 1) planning, measuring, and qualifying parts and 2) enhancing quality through a processing-monitoring-qualifying framework. Creating this new system requires building the complex relationships between requirements, processes, and quality. These relationships are needed to specify, measure, analyze, and optimize variables to ensure final part quality. Thus, a processing-monitoring-quality framework could provide critical steps to identify those relationships and help meet stakeholder needs. The paper describes how to adapt the “software and systems engineering” V-model to a “metal AM quality assurance” V-model that can provide a framework for quality assurance in MAM.

This paper introduces a graphical representation based on the fusion of several disparate standards to instantiate a sensor wrapper and sensing schema for fault delineation in machine tools and other manufacturing assets. Texas A&M researchers have already developed a sensor wrapper that aims to specify the sensor and the sensing suite based on a systematic consideration of the functionality (based on process dynamics) to derive the configuration and instantiation of a viable sensing suite. Adapting this scheme for real-world machines and manufacturing assets is challenging because of the complexity of the machine tool structure and the diversity of faults within its components. The presented graphical representation method is based on an ontological compliance with MTConnect and International Organization for Standardization/International Electrotechnical Commission standards, and the representation employs the graphical motifs pertaining to the fault tree framework. Such representation is essential for the delineation of failure modes associated with components of a machine tool, thereby making sensor-wrappers viable for the smartification of machine tools. The issues pruning the levels of the graphical representation based on the domain knowledge of the machine tool, and the many-to-many mapping between the components and sensors are discussed. The representation was applied in order to derive suitable sensing schemes for conventional machine tools, i.e., lathe and milling machines and a modern hybrid additive manufacturing machine.

Additive manufacturing (AM) provides a higher level of flexibility to build customized products with complex geometries, by selectively melting and solidifying metal powders. However, wide applications of AM beyond rapid prototyping are currently limited by its ability to perform quality assurance and control. Advanced melt-pool monitoring provides a unique opportunity to increase information visibility in the AM process. Stochastic melt-pool variations are closely pertinent to the quality of an AM build. There is a pressing need to investigate the variances of melt pools along the temporal scanning path, as well as within a 3D spatial neighborhood of the focal point by the laser beam. This paper presents a stochastic modeling framework to characterize and monitor spatiotemporal variations of melt-pool imaging data, including tensor decomposition of high-dimensional data, additive Gaussian process modeling of low-dimensional profiles as random variables, and hypothesis testing via the construction of confidence boundary for statistical process monitoring. Experimental results show the effectiveness of tensor decomposition for spatiotemporal monitoring of melt-pool variations in the metal-based AM process.

In a high‐mix and low‐volume manufacturing facility, heterogeneous jobs introduce frequent reconfiguration of machines which increases the chance of unplanned machine breakdowns. As machines are often nonidentical and their performance degrades over time, it is critical to consider the heterogeneity and non‐stationarity of the machines during scheduling. We propose a reinforcement learning‐based framework with a novel sampling method to train the agent to schedule heterogeneous jobs on non‐stationary unreliable parallel machines to minimize weighted tardiness. The results indicate that the new sampling approach expedites the learning process and the resulting policy significantly outperforms static dispatching rules.

Healthcare experts have come to a consensus that effective and safe vaccines are necessary to control the rapid spread of the ongoing COVID-19 pandemic across the globe. Since the traditional vaccine development and manufacturing approaches were unable to meet the rapidly growing COVID-19 vaccine demand, biopharmaceutical firms had to devise novel and smart techniques to boost the development, production, and distribution of COVID-19 vaccines in a large scale with lightning speed. This triggered their transition to smart vaccine manufacturing approaches using novel viral vector and nucleic acid biotechnology platforms. This paper tries to explore this rationality of the biopharmaceutical industry by comparing the traditional and the novel biotechnology platform-based vaccine manufacturing techniques and reviewing the COVID-19 vaccine manufacturing scenarios. To highlight the “smart” characteristics of the novel platform-based COVID-19 vaccine products and to make an effective comparison with the traditional products, a well-established product classification framework is used as a reference. Finally, the study concludes by presenting the future possibility of incorporating smart manufacturing paradigms with the novel platform-based manufacturing process. It is hoped that this study would serve as an asset for the biopharmaceutical firms to appropriately streamline their strategies, resources, and goals to meet the global vaccine requirements.