Publications

Accurate phase identification in large-scale molecular dynamics simulation remains a significant challenge due to ambiguous boundaries between compositionally distinct regions and the lack of ground truth labels. While unsupervised methods can perform phase segmentation for small systems through structure-aware segmentation pipelines, their computational cost becomes prohibitive for large-scale analysis. We present a weaklysupervised machine learning pipeline that trains Graph Neural Networks (GNNs) to enable scalable phase segmentation in 3D atomistic systems. Using a physically grounded unsupervised method, we generate weak labels for small FeMgSiON systems that exhibit Fe-rich (metallic) and Fe-poor (silicate) phase separation. These labels guide GNNs to learn physically meaningful representations of atomic neighborhoods. Once trained, the GNNs act as an efficient parametric model, enabling direct segmentation of arbitrarily large atomistic systems eliminating the computational overhead of the initial unsupervised pipeline. By learning from thousands of weakly labeled snapshots, the model discerns latent structural patterns, enhancing both prediction accuracy and generalization to unseen data. This methodology enables efficient, accurate, and physically consistent phase segmentation in large-scale molecular dynamics, unlocking new possibilities for scalable analysis in material simulations.

Nitrogen as the dominant volatile element of the atmosphere is also expected to exist underneath the Earth’s surface. Its interior budget and distribution may have largely been set early on by the core formation and other processes that led to apparent volatile depletion in the terrestrial planet. To better understand how nitrogen may have behaved during the accretion stages when proto-Earth experienced multiple episodes of magma ocean environments, we report a first-principles computational study of nitrogen incorporation in silicate melts over wide pressure range 0 to 125 GPa (3000 to 5000 K), as well as in iron-rich metallic melt and bulk Earth-like melt at selected conditions. The results show that the speciation of nitrogen in silicate melts at low pressures consists of almost entirely N2 molecules with interstitial occupancy. As pressure increases, nitrogen interacts increasingly with the silicate network and bonds with iron more strongly than with any other cations (Mg, Si, Ca, Al, and Ni) present in the melt. Both pressure and reducing conditions help nitrogen chemically dissolve as nitride species thus promoting nitrogen retention of possible deep-seated dense silicate melts in the mantle. Metallic liquid incorporates nitrogen by bonding with iron with weak or no interactions with itself or with other impurities. The simulated bulk Earth melt system shows a phase segregation to form an iron-rich cluster which is surrounded by a silicate region. The metal-silicate partition coefficient of nitrogen evaluated using the relevant local coordination statistics from the two-phase system is ∼31 at 30.5 GPa (3000 K, IW-3.1), ∼18 at 37.1 GPa (4000 K, IW-2.2), and ∼24 at 131 GPa (5000 K, IW-2.1) which are generally consistent with the measured trends. Based on the predicted strong preferential partitioning to the metallic liquid, we argue that while nitrogen may be depleted from the silicate mantle, it may be sequestered in the core.

The earth is thought to have gone through complex physicochemical changes during the accretion and magma ocean stages. To better understand this evolution process at the fundamental level, we investigate the behavior of a bulk earth melt system by simulating the composition Fe35.7Mg19.0Si15.2O30.2 (in wt%) at high pressure. A deep neural network potential trained by first-principles data can enable accurate molecular dynamics simulation of large supercells that greatly enhances sampling for reliable evaluation of elemental partitioning. Our simulated system undergoes a phase separation in which the four elements clump together to different extents into two major domains. Based on the coordination and space-decomposition analyses, the inferred composition at 3000 K and 29.1 GPa contains 96.2, 0.1, 1.9 and 1.7 wt% of Fe, Mg, Si, and O, respectively, for the one domain and the corresponding elemental proportions are 3.0, 29.7, 22.0, and 45.3 wt% for the other domain. The predicted segregation thus leads to the formation of an iron-rich phase which corresponds to the metallic core and a magma ocean phase which corresponds to the silicate mantle. The metallic domain incorporates more silicon and more oxygen whereas the magma ocean domain gains more iron oxides at higher temperatures. Our predicted compositions compare favorably with those derived from experimental work for the equilibrium state metal and silicate reacting under high-pressure conditions.

Stock price prediction has been a trending yet mystifying topic for a very long time now. The fact that the stock price on short-term are the result of public reaction to rumours and are moreover, associated with public psychology, has made it tricky for the general people to get an insight on the market. This paper presents a method to predict stock prices of companies listed under Nepal Stock Exchange Limited at an interval of every two minutes. The paper describes an approach to comprehend instant public preferences through traded share volume, number of transactions and price fluctuation analysis. The result of these analysis are fed to Neural Network, which then predicts the percentage change in stock price.