Elemental abundances, which are often depleted with respect to solar values, are important input parameters for kinetic models of interstellar chemistry. In particular, the amount of carbon relative to oxygen is known to have a strong effect on the modeled abundances of many species. While previous studies have focused on comparison of modeled and observed abundances to constrain the C/O ratio, the effects of this parameter on the underlying chemistry have not been well studied. We investigated the role of the C/O ratio on dark-cloud chemistry using the NAUTILUS code and machine learning techniques for molecular representation. We find that modeled abundances are quite sensitive to the C/O ratio, especially for carbon-rich species such as carbon chains and polycyclic aromatic hydrocarbons. CO and simple ice-phase species are found to be major carbon reservoirs under both oxygen-poor and oxygen-rich conditions. The appearance of C _3 H _4 isomers as significant carbon reservoirs, even under oxygen-rich conditions, indicates the efficiency of gas-phase C _3 formation followed by adsorption and grain-surface hydrogenation. Our model is not able to reproduce the observed, gas-phase C/H ratio of TMC-1 CP at the time of best fit with any C/O ratio between 0.1 and 3, suggesting that the modeled freeze-out of carbon-bearing molecules may be too rapid. Future investigations are needed to understand the reactivity of major carbon reservoirs and their conversion to complex organic molecules.
Sulfur dioxide (SO _2 ) ice has been tentatively detected in protostellar envelopes, but its reliability as a solid-state sulfur reservoir remains unclear. We present new measurements of SO _2 ice from 6.8–8.5 μ m toward four Class 0 protostars observed with JWST’s Mid-Infrared (MIRI) Medium Resolution Spectrometer, as part of the COMs ORigin Investigated by the Next-generation Observatory in Space (CORINOS) program. The sample spans a luminosity range from 1 L _⊙ (B335, IRAS 15398–3359) to 10 L _⊙ (L483, Ser-emb 7). To assess continuum placement uncertainty in absorption spectra, we apply randomized polynomial fits over the restricted region. We fit laboratory spectra from the Leiden Ice Database for Astrochemistry using the open-source Python library Omnifit. We detect the 7.7 μ m CH _4 band in all sources and find its column density robust to baseline choice, providing a reference for evaluating the weaker SO _2 feature on its blue shoulder and quantifying baseline-related uncertainty. Three SO _2 laboratory ices were tested: pure SO _2 ice yields 0.3%–1.2% of volatile sulfur may be locked in SO _2 ice (lower and upper limits); CH _3 OH:SO _2 ice gives 0.02%–0.18%, but with lower quality fitting. The best-fitting H _2 O:SO _2 ice yields 0.2%–0.9%, which we consider the most realistic. These ranges define plausible bounds on SO _2 ice abundances in our sample. We find evidence for SO _2 in Ser-emb 7, L483, and IRAS 15398–3359, but emphasize the noisy spectrum of B335 prevents a definitive detection. Comparing SO _2 ice abundances across the different environments, we assess how conditions influence the role of SO _2 as a potential sulfur reservoir and implications for the longstanding “missing sulfur” problem.
The increase in the number of phishing and scam websites adds a continuous threat to online users. In this research, we present TrustNet, a hybrid AI platform that integrates traditional machine learning (ML) models with large language models (LLMs). We used different machine learning algorithms, and the Random Forest classifier achieved the highest performance, with an Accuracy of 0.9694, a Precision of 0.9749, a Recall of 0.9628, and an F1 of 0.9688. Further feature correlation and network centrality analyses showed the most influential attributes and their contribution to characterizing scam websites. Additionally, we built a multi-agent LLM system to identify and analyze linguistic and contextual indicators of scams in website content. The system contains four specialized agents (Market Researcher, Language Analyst, Threat Analyst, and Evaluator). The agents work together to identify urgency, vagueness, grammatical inconsistencies, and deceptive pressure tactics in web content. We tested the LLM-based agent system on our custom dataset of real and synthetic website text samples. The team of agents achieved high detection performance with an Accuracy of 0.9931, a Precision of 0.9867, a Recall of 1.0, and an F1 score of 0.9933. Both ML models and LLM-driven agents were integrated into a web-based application named TrustNet. TrustNet is a platform for detecting scams and phishing, and for explainable risk assessment. The system provides an interactive dashboard with risk-level visualization, linguistic explanation panels, and real-time recommendations to users. The proposed hybrid framework advances the state of the art in scam website detection by using both ML and LLM in a mixed approach that conducts both quantitative URL analysis and qualitative linguistic, contextual, and semantic knowledge extraction.
Non-premixed turbulent combustion is central to many practical energy conversion systems, where turbulence drives the mixing of the fuel and oxidizer at the molecular scale in high-Reynolds number flows. In this study, high-order compact finite-difference schemes are studied as a means of enhancing the numerical fidelity of steady flamelet-based large eddy simulation (LES) for turbulent non-premixed flames. A Fortran-based LES model employing high-order compact finite-difference discretization is developed within a steady flamelet framework to solve the filtered compressible Navier–Stokes equations with an implicit Smagorinsky–Lilly subgrid-scale model in cylindrical coordinates. Spatial discretization is performed using fifth- and fourth-order compact schemes for the convective and viscous terms, respectively, while temporal integration is carried out using a fourth-order Runge–Kutta method. The computational framework is validated using benchmark turbulent pipe-flow data, along with detailed experimental measurements from the Sandia piloted Flame D benchmark flame. The proposed LES framework provides statistically converged predictions of velocity, mixture fraction, temperature, and major species (H2O and CO2). Moderate discrepancies are observed for carbon monoxide (CO), which are attributed to the steady-state chemistry assumption and its inability to fully represent finite-rate oxidation processes in post-flame regions. Predictions of nitric oxide (NO) show a marked improvement compared to results obtained using a conventional flamelet/progress-variable (FPV) formulation, highlighting the benefit of enhanced numerical fidelity in reducing numerical dissipation within a steady flamelet framework. Comparisons with FPV formulations that incorporate a separate transport equation for NO suggest that explicitly accounting for the transient evolution of NO remains important for accurately capturing the behavior of slow-forming species. The results suggest that the use of high-order compact finite-difference schemes presents coherent numerical formulation with enhanced accuracy compared to traditional FPV formulation in the simulation of turbulent flames and pollutant formation.
Effective cybersecurity training requires preparing personnel at every organizational level to operate under pressure, interpret incomplete information, and make rapid, high-stakes decisions. However, current training practices struggle to capture the dynamic, adaptive, and adversarial nature of real-world incidents. These methods lack realism, rely on rigid scripts, and fail to model evolving attacker behavior or the cascading consequences of defensive decisions. As a result, organizations continue to show substantial readiness gaps. CISA's recent assessments indicate that most victims remain unprepared across both technical controls and decision-making processes, from junior analysts to executive leadership. Meanwhile, large-scale exercises such as Cyber Storm demonstrate the value of adaptive, inject-driven training but demand resources and coordination not feasible for routine use. This gap motivates the need for a training system that combines the accessibility of tabletop exercises with the realism and adaptivity of full-scale cyber simulations. In this work, we introduce an AI-driven immersive emulation platform that transforms traditional tabletop scenarios into dynamic, decision-responsive training experiences. The system integrates multiple specialized AI agents (Including adversary emulation, scenario orchestration, content generation, and performance analysis) coordinated through a modular Model Context Protocol (MCP) architecture. These components enable real-time scenario evolution, realistic attack behaviors aligned with the MITRE ATT&CK framework, and automated generation of supporting artifacts such as logs, emails, and incident injects. The platform also delivers structured, role-specific feedback through a performance intelligence engine that evaluates both quantitative metrics and qualitative decision-making. By blending adaptive AI agents with scalable microservice architecture, the system offers a practical, repeatable, and highly realistic alternative to traditional tabletop training. We expect this approach to measurably improve situational awareness, decision-making speed, ethical and legal compliance, communication effectiveness, and overall readiness during cybersecurity incidents.
