Bourns College of Engineering

The rapid growth of generative AI has placed significant strain on traditional data center infrastructures and existing power grids, leading to soaring energy demands and environmental burdens that may disproportionately affect the local communities. Shifting AI inference from the cloud to edge devices could potentially reduce the reliance on network connections, enhance user privacy, and alleviate the escalating pressure data centers impose on the local electricity grid. In this work, we present a case study examining the environmental footprint and energy consumption when deploying a generative AI model on cloud and edge platforms. To this end, we model and evaluate the water consumption and carbon emissions associated with AI inference across these deployment scenarios. Our empirical results demonstrate that, for several state-of-the-art generative AI models deployable on both cloud and edge devices, a reduced environmental footprint is observed for edge platform deployments. More specifically, edge platforms can achieve over 90% energy savings while reducing carbon emissions and water consumption by more than 80%. Putting the accuracy and latency performance aside, these findings highlight the potential of edge inference to lower the energy demands and environmental footprint of generative AI compared to cloud-based inference.

In this work, we developed a dual-mode portable device that integrated a 3D-printed microfluidic chip for detection of dipicolinic acid (DPA) on chip. The system uses a ratiometric fluorescence nanoprobe formed by embedding carbon dots (CDs) into an Eu3⁺ metal–organic framework (Eu-MOF). Upon reaction with DPA in the microchannel, red fluorescence was enhanced and blue fluorescence suppressed, enabling sensitive ratiometric detection of DPA on chip with a detection limit (LOD) of 0.04 µM. Interestingly, the composite EuMOF/CDs/DPA also exhibits peroxidase-like activity, catalyzing the oxidation of TMB into a blue-colored product (oxTMB), which allows for colorimetric detection with an LOD of 10.14 µM. To improve usability and reduce environmental or instrumental variability, incorporating a microfluidic chip into a semi-portable device and utilizing a smartphone, making the system portable and miniaturized for easy operation. In the smartphone-assisted mode, the LODs were 0.33 µM (ratiometric fluorescence) and 12.27 µM (colorimetry), determined by RGB signal analysis, respectively. Moreover, satisfactory recoveries (85–104.6%) were achieved in the spiked real samples. Overall, this platform offers a straightforward, cost-effective, and versatile approach for DPA detection, with promising applications in food safety, environmental monitoring, and clinical diagnostics.

Computational cognitive models are powerful tools for enhancing the quantitative and theoretical rigor of cognitive neuroscience. It is thus imperative that model users-researchers who develop models, use existing models, or integrate model-based findings into their own research-understand how these tools work and what factors need to be considered when engaging with them. To this end, we developed a philosophical toolkit that addresses core questions about computational cognitive models in the brain and behavioral sciences. Drawing on recent advances in the philosophy of modeling, we highlight the central role of model users' reasoning goals in the application and interpretation of formal models. We demonstrate the utility of this perspective by first offering a philosophical introduction to the highly popular drift diffusion model (DDM) and then providing a novel conceptual analysis of a long-standing debate about decision thresholds in the DDM. Contrary to most existing work, we suggest that the two model structures implicated in the debate offer complementary-rather than competing-explanations of speeded choice behavior. Further, we show how the type of explanation provided by each form of the model (parsimonious and normative) reflects the reasoning goals of the communities of users who developed them (cognitive psychometricians and theoretical decision scientists, respectively). We conclude our analysis by offering readers a principled heuristic for deciding which of the models to use, thus concretely demonstrating the conceptual and practical utility of philosophy for resolving meta-scientific challenges in the brain and behavioral sciences.

Genetic linkage maps are an essential tool in population genetics and plant breeding research, yet user-friendly online tools for constructing and visualizing them remain scarce. MSTmap Online addresses this gap by providing a modern, accessible platform for generating high-quality genetic linkage maps from genotypic data. The web server quickly computes linkage groups using the MSTmap algorithm and generates detailed output files, including publication-ready PDF visualizations of linkage groups. The server supports bookmarking and asynchronous processing, allowing users to revisit their results at a later time. A companion Python library for MSTmap Online enables seamless integration into custom analysis pipelines. MSTmap Online is free and open to all users with no login requirement at https://mstmap.org. The companion Python library is available at https://pypi.org/project/mstmap/.

The discovery of compressed atomic-type hydrides offers a promising avenue toward achieving room-temperature superconductivity, but it necessitates extremely high pressures to completely dissociate hydrogen molecules to release free electrons. Here, we report a remarkable finding of compressed H2-molecular-type hydride CaH14 exhibiting an unusual transition temperature (Tc) of 204.0 kelvin. The peculiarity of its electronic structure lies in the pronounced emergence of near-free electrons, which manifest metallic bonding, but molecular hydrogen fragments persist. This finding indicates that the necessary condition for superconducting transition is forming the Fermi sea with Cooper pairs rather than the monatomic hydrogen. Notably, the formation mechanism of free electrons can be effectively explained by the finite-depth potential wells model. Intriguingly, this H2-molecular-type hydride can downgrade the required pressure to 80 gigapascal while maintaining a high Tc of 84 kelvin, well above the liquid-nitrogen temperature. Our study has established a high-temperature superconducting paradigm and opened the prospect for achieving high-Tc superconductors in H2-molecular-type hydrides at low pressure.

Fires in the wildland-urban interface (WUI) are a global issue with growing importance. However, the impact of WUI fires on air quality and health is less understood compared to that of fires in wildland. We analyze WUI fire impacts on air quality and health at the global scale using a multi-scale atmospheric chemistry model-the Multi-Scale Infrastructure for Chemistry and Aerosols model (MUSICA). WUI fires have notable impacts on key air pollutants [e.g., carbon monoxide (CO), nitrogen dioxide (NO2), fine particulate matter (PM2.5), and ozone (O3)]. The health impact of WUI fire emission is disproportionately large compared to wildland fires primarily because WUI fires are closer to human settlement. Globally, the fraction of WUI fire-caused annual premature deaths (APDs) to all fire-caused APDs is about three times of the fraction of WUI fire emissions to all fire emissions. The developed model framework can be applied to address critical needs in understanding and mitigating WUI fires and their impacts.

The hydrogenation of CO2 to CH3OH over Cu-based catalysts holds significant potential for advancing carbon sequestration and sustainable chemical processes. While numerous studies have focused on catalyst development, the environmental effects on underlying reaction mechanisms have yet to be fully understood. In this work, we develop a grand potential theory for a comprehensive analysis of CO2 hydrogenation to CH3OH over Cu (111) and Cu (211) surfaces. By integrating electronic and classical density functional calculations to bridge the pressure gap, the theoretical results revealed that the HCOO* formation rate may vary by several orders of magnitude depending on reaction conditions. The grand potential theory enables us to elucidate the molecular mechanisms underlying the need for high H2 pressure, the prevalence of saturated CO2 adsorption, and the important roles of CO and H2O in hydrogenation. Moreover, this study addressed and clarified controversies over CO2 versus CO adsorption and hydrogenation, the formate versus carboxy pathways, and the difference in HCOO* hydrogenation activity between Cu (111) and Cu (211) surfaces. The theoretical analysis offers a new perspective for optimizing reaction conditions and catalyst performance in methanol synthesis and can be generalized to enhance our understanding of heterogeneous catalysis under industrially relevant conditions.

DNA methylation is an epigenetic marker that directly or indirectly regulates several critical cellular processes. While cytosines in mammalian genomes generally maintain stable methylation patterns over time, other cytosines that belong to specific regulatory regions, such as promoters and enhancers, can exhibit dynamic changes. These changes in methylation are driven by a complex cellular machinery, in which the enzymes DNMT3 and TET play key roles. The objective of this study is to design a machine learning model capable of accurately predicting which cytosines have a fluctuating methylation level [hereafter called differentially methylated cytosines (DMCs)] from the surrounding DNA sequence. Here, we introduce L-MAP, a transformer-based large language model that is trained on DNMT3-knockout and TET-knockout data in human and mouse embryonic stem cells. Our extensive experimental results demonstrate the high accuracy of L-MAP in predicting DMCs. Our experiments also explore whether a classifier trained on human knockout data could predict DMCs in the mouse genome (and vice versa), and whether a classifier trained on DNMT3 knockout data could predict DMCs in TET knockouts (and vice versa). L-MAP enables the identification of sequence motifs associated with the enzymatic activity of DNMT3 and TET, which include known motifs but also novel binding sites that could provide new insights into DNA methylation in stem cells. L-MAP is available at https://github.com/ucrbioinfo/dmc_prediction.

Repetitive regions in eukaryotic genomes often contain important functional or regulatory elements. Despite significant algorithmic and technological advancements in genome sequencing and assembly over the past three decades, modern de novo assemblers still struggle to accurately reconstruct highly repetitive regions. In this work, we introduce RAmbler (Repeat Assembler), a reference-guided assembler specialized for the assembly of complex repetitive regions exclusively from PacBio HiFi reads. RAmbler (i) identifies repetitive regions by detecting unusually high coverage regions after mapping HiFi reads to the draft genome assembly, (ii) finds single-copy k-mers from the HiFi reads, (i.e., k-mers that are expected to occur only once in the genome), (iii) uses the relative location of single-copy k-mers to barcode each HiFi read, (iv) clusters HiFi reads based on their shared bar-codes, (v) generates contigs by assembling the reads in each cluster, and (vi) generates a consensus assembly from the overlap graph of the assembled contigs. Here we show that RAmbler can reconstruct human centromeres and other complex repeats to a quality comparable to the manually-curated telomere-to-telomere human genome assembly. Across over 250 synthetic datasets, RAmbler outperforms hifiasm, LJA, HiCANU, and Verkko across various parameters such as repeat lengths, number of repeats, heterozygosity rates and depth of sequencing.

Cancer, a significant cause of mortality, necessitates improved drug delivery strategies. Exosomes, as natural drug carriers, offer a more efficient, targeted, and less toxic drug delivery system compared to direct dispersal methods via ingestion or injection. To be successfully implemented as drug carriers, efficient loading of drugs into exosomes is crucial, and a deeper understanding of the loading mechanism remains to be solved. This study introduces surface-enhanced Raman scattering (SERS) to monitor drug loading efficacy at the single vesicle level. By enhancing the Raman signal, SERS overcomes limitations in Raman spectroscopy. A gold nanopyramids array-based SERS substrate assesses exosome heterogeneity in drug-loading capabilities with the help of single-layer graphene for precise quantification. This research advances targeted drug delivery by presenting a more efficient method of evaluating drug-loading efficiency into individual exosomes through SERS-based monitoring. Furthermore, the study explores leveraging osmotic pressure variations, enhancing the efficiency of drug loading into exosomes.