At its core, MyQL utilizes retrieval-augmented generation, a technique that allows the AI to ground its responses in specific, verifiable sources rather than generating content from general probability models. Instructors can upload relevant course materials, such as lecture notes, textbooks, or syllabi, which the system then uses to generate context-aware answers. Every response provided to the student includes citations pointing back to the specific source material. This feature ensures that the information students receive is precise and directly related to their coursework, reducing the risk of hallucinations or irrelevant information that can occur with less specialized AI tools.

A distinctive feature of the MyQL platform is its emphasis on fostering critical thinking through scaffolded learning. The system does not merely provide correct answers; it actively encourages students to evaluate and correct information. To achieve this, the platform can introduce intentional inaccuracies into its responses or practice questions. Students are then prompted to assess the validity of the information and identify errors. This method promotes active engagement and helps learners develop the skills necessary to work effectively alongside AI systems, rather than passively consuming the output they receive.

The platform is structured around three core components: the integration of course material, interactive learning, and experiential learning. The interactive component includes practice questions and tailored feedback that guide students step-by-step through the material. The experiential learning aspect involves personalized assignments focused on key learning goals. In these scenarios, the AI manages secondary tasks, allowing students to focus on decision-making and practical application. This setup enables students to observe the implications of their choices in real-time, adapting their strategies as they progress through the coursework.

MyQL is designed with accessibility and privacy as primary considerations. The technology is optimized to operate efficiently on low-end hardware, making it suitable for diverse educational environments where high-performance computing resources may be limited. By prioritizing user privacy, the platform aims to safeguard student data and ensure that the learning experience remains secure. The system maintains a high degree of originality in its code, which reduces reliance on third-party content and enhances security and customization options for educational institutions.

The development of MyQL reflects a broader need for tools that support both instructors and students in an era of increasing AI integration. Traditional educational methods often struggle to maintain student engagement when AI tools are used primarily for information retrieval. MyQL addresses this by creating an environment where AI augments the teaching process rather than replacing it. The platform supports collaborative learning by facilitating interactions between students and the AI, helping to build skills that are increasingly relevant in the modern workforce.

Potential applications for MyQL are extensive, spanning higher education institutions, online and hybrid learning environments, and workforce training programs. Schools aiming to train students in collaborative skills with emerging AI technologies may find the platform particularly useful. Additionally, educational programs focused on experiential learning and real-time problem-solving can leverage the platform’s ability to simulate complex scenarios. The technology is also relevant for settings where low-cost, privacy-conscious AI tools are essential due to limited hardware resources, such as community colleges or underfunded schools.

Currently, the technology is at Technology Readiness Level 4, indicating that it has been validated in a laboratory environment. The intellectual property status is patent pending, and the Research Foundation for the State University of New York has indicated that the technology is available for licensing. This stage of development suggests that while the core functionality has been demonstrated, further testing and refinement may be required before widespread deployment in live educational settings. The availability for licensing allows institutions to adapt the tool to their specific needs, ensuring that it aligns with their pedagogical goals and technical infrastructure.

The announcement of MyQL highlights the ongoing efforts by public university systems to innovate in the field of educational technology. By leveraging AI to enhance, rather than replace, the human element of teaching, MyQL represents a step toward more personalized and effective learning experiences. As AI continues to evolve, tools that prioritize accuracy, critical thinking, and accessibility will likely play a crucial role in shaping the future of education. The platform’s focus on integrating course materials and promoting active learning positions it as a significant development in the field of AI-augmented tutoring.

Subject of Research: Education

Article Title: MyQL: the retrieval augmented tutor

Article References: MyQL: the retrieval augmented tutor. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: Artificial Intelligence, Educational Technology, Retrieval-Augmented Generation, SUNY, Online Learning, Critical Thinking, AI Tutoring, MyQL, retrieval, augmented, tutor, scientific research

The central insight behind QEGuard is that a previous calculation can be relevant without being transferable. Two records may be similar enough to invite comparison, yet differ in conditions that determine whether specific parameters can actually be reused. Switching from a primitive to a conventional unit cell, for example, makes direct reuse of k-point settings inappropriate, because the same reciprocal-space resolution is expressed through different grid indices. A density-of-states calculation may contain occupation settings unsuitable for a baseline self-consistent-field run, and a non-converged calculation can remain structurally informative while being entirely unsuitable as a source of defaults. In current practice, researchers embed these judgments in implicit rules, such as borrowing from the most similar successful run, and such rules conflate decisions with fundamentally different transfer semantics.

QEGuard implements a staged architecture that separates record relevance from transfer eligibility. First, a parser converts Quantum ESPRESSO stdout and XML outputs into canonical JSON-compatible records containing structural descriptors, calculation settings, workflow context, result summaries, and quality signals. The parser extracts cutoff values, k-point grids, occupation and smearing settings, total and Fermi energies, band-gap metadata, cell vectors, atomic positions, and symmetry descriptors, using the Atomic Simulation Environment for structure handling and spglib for symmetry analysis. Second, a retrieval layer stores these records in a SQLite-backed database and ranks candidate precedents using a structure-aware score that combines formula agreement, cell-scale geometry, local-environment proxies, symmetry signals, and compact fingerprints.

That retrieval score proved remarkably effective in benchmarking. Across 3352 queries drawn from 5777 JARVIS-derived records, the structure-aware score achieved 83.7 percent Top-1 success, meaning it recovered the intended source record as the highest-ranked candidate, compared with 48.3 percent for formula-only retrieval, 76.1 percent for formula-gated pymatgen StructureMatcher, and 78.4 percent for SOAP descriptors. The authors stress that this score is designed for candidate selection within their parsed-record setting rather than as a universal crystal-similarity metric, and its role is simply to supply relevant precedents to the guarded transfer stage.

The transfer-decision component is where QEGuard departs most clearly from conventional practice. It evaluates each retrieved candidate against the target calculation context using method compatibility, scientific-objective agreement, accuracy-profile sufficiency, candidate quality, representation relationships, workflow-context matching, and parameter-level transfer rules. The output is a structured decision record specifying which parameters can be proposed for reuse, which are excluded, and the reason for each exclusion. Three outcomes are possible: a full transfer when all checks pass, a partial transfer applying only eligible fields, or an abstention that returns no precedent-derived updates when no candidate qualifies. Notably, abstention is treated as a first-class policy outcome rather than a retrieval failure.

When a cell representation changes, the package can derive an initial k-point mesh by preserving reciprocal-space resolution rather than blindly copying cell-index grids. In validation tests, directly reusing a conventional-cell 7x7x7 mesh for silicon in its primitive cell produced a 4.67 meV/atom energy difference, whereas the resolution-preserving approach generated a 12x12x12 mesh with a difference of only 0.023 meV/atom. For molybdenum disulfide, mapping a 6x6x1 primitive-cell mesh to a 3x3x1 supercell mesh reduced observed wall time by a factor of 1.84, and for anatase titanium dioxide the mapped mesh reduced wall time by a factor of 2.65 while matching an independent reference to within 0.000295 meV/atom. These derived meshes remain starting values subject to target-specific convergence validation.

The negative-control experiments illustrate why such safeguards matter. Across nine deliberately incompatible cases, unguarded top-ranked reuse transferred at least one ineligible field in every case, propagating 27 of 39 predefined ineligible parameters. QEGuard transferred none of them while retaining all nine predefined eligible fields. One striking negative control involved a self-consistent-field calculation that printed JOB DONE and returned process code zero despite reporting non-convergence; a process-completion criterion would have exposed its parameters for reuse, but QEGuard rejected the record based on the parsed convergence state. Another showed that screening settings with errors of 20.22 and 74.64 meV/atom were being applied to a 1 meV/atom accuracy target, a mismatch the software correctly refused.

Controlled evaluation reinforced these findings. Across 74 deterministic policy cases spanning the titanium dioxide and silica/zirconia families, covering compatible precedents, polymorph discrimination, workflow-conditioned ranking, representation changes, quality criteria, and compatibility metadata, QEGuard achieved perfect Top-1 source recovery, correct transfer, and zero unsafe transfer. A titanium dioxide case further demonstrated that parameter adequacy depends on the scientific objective: at a 37.5/300 Ry cutoff the energy error reached 74.64 meV/atom, yet the maximum force-component error of 0.00222 eV/Å comfortably satisfied the force criterion, showing that the same settings can be inadequate for one property and acceptable for another.

QEGuard complements rather than competes with existing infrastructure. Workflow systems such as AiiDA, FireWorks, and Atomate provide provenance capture and high-throughput execution, while Materials Cloud and NOMAD support dissemination and reuse of simulation data, but none addresses the narrower question of which portion of an available prior record can responsibly inform a new calculation. Similarly, community parameter templates offer general defaults, whereas QEGuard evaluates whether a provenance-bearing local record can supply context-specific starting parameters for a declared objective and accuracy requirement. The package occupies the gap between record availability and authoring-time parameter reuse.

The software is released under the MIT license, version 0.1.0 is available on GitHub, and the repository includes reproducible examples, schema documentation, benchmark data, and resources for reproducing the numerical-validation calculations. Although the current implementation is specific to Quantum ESPRESSO, the staged interface is designed to support other density functional theory codes through additional parsers, parameter mappings, and code-specific validation of transfer rules. The authors argue that negative decisions are themselves valuable output: exclusion can be as informative as transfer, and structured decision records allow downstream tools to retain both proposed fields and the reasons for withholding others, making parameter reuse explicit, inspectable, and reproducible.

Subject of Research: Guarded reuse of parameters from prior Quantum ESPRESSO density functional theory simulation records

Article Title: QEGuard: A python package for guarded precedent reuse in Quantum ESPRESSO simulation records

Article References: Kim, H., Kim, H., Han, S., & Lee, J. (2026). QEGuard: A python package for guarded precedent reuse in Quantum ESPRESSO simulation records. SoftwareX, 36, Article 103017. https://doi.org/10.1016/j.softx.2026.103017

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103017

Keywords: QEGuard, Quantum ESPRESSO, density functional theory, computational materials science, parameter reuse, k-point mesh, simulation records, retrieval, SQLite, pseudopotentials, workflow systems, open-source software