cs.AIOct 13, 2026
SaveComplex materials engineering via Provenance Action Graphs: potential, structure and demonstration
Organizations: Department of Materials Engineering, Ben Gurion University of the Negev, Beer-Sheva, Israel · Department of Chemistry, Colorado State University, Fort Collins, CO 80523 · Department Materials, University of California, Santa Barbara, Santa Barbara, CA 93106
Abstract
This work discusses and tests the potential of reaction graphs for material synthesis, in which events consist of reaction steps (actions) and their outputs (observations). The importance comes from the requirement to synthesize technologically useful materials, especially those that are kinetically stabilized. Often, it is a science that follows years of trial and error, culminating in (published) successful recipes, but is truly driven by far more (unpublished) failures. The entry into the era of AI provides an opportunity to use learning algorithms to encode rules from multi-dimensional synthesis pathways, regardless of their technological usefulness, and to decode both forward prediction and inverse learning modes. The main incentive is to accelerate material discovery by generating a global knowledge base that can be shared among data agents and contributed to by them. We will focus on several scenarios in the common synthesis process during solid-state synthesis, primarily because these reactions are kinetic in nature and require closer monitoring and control of the reaction history.
Figures & tables
Figure 1: Basic principles of a MPG .(a) Illustrations of a synthesis process that is represented as a time series of event-nodes: actions and observables, which may carry attributes in a form of configurations. (b) Representation of a MPG in a way that emphasize its properties: actions as nodes containing information regarding the different operations that happen during a synthesis (‘add’, ‘mix’, ‘wait’), which are inter-connected by edges that describe their sequnce-related topology; observations as sink-nodes containing information regarding the outcome of a sequence of actions (‘color’, ‘shape’, ‘structure’), and that happen at a particular time between actions; configurations are attributes to action or observation nodes, which carry unique information about the event-node. The different colors emphasize the idea of granularity of a MPG , meaning that a MPG can be represented as a sequence of sub-graphs.
Figure 2: Data structure and internal relationships : (a) The central material’s descriptors for rationalizing potential Adaptivity. “History” refers to events that occur in sequence and include actions and environmental attributes. “State” refers to descriptors on which a material’s property depends (e.g., chemistry, local order, density, and their spatial distribution), and a dependent variable of History. “Property” refers to a function-critical feature of the material (e.g., conductance, permittivity), and a dependent variable of State. (b) Criteria for “Adaptivity” conditions, showing the required relationship between History, States, and Properties. (c) A strategy for interacting and navigating in the multi-dimensional space, beyond manual exploration, using self-supervised autoencoder-based projected latent spaces. The projected latent spaces enable critical dimensionality reduction to identify correlations among different descriptors. Selection of an area/volume in latent space can allow back-projection to the original data space between the descriptor’s projected latent spaces; this requires a common dimension; here, ‘ n ’ as the number of experiments. (d) Generating a correlative model that allows interpolation between descriptors will be done using ML transformers. (e) Priming physical states and targeting properties will be done, for example, using Model-based Reinforcement learning to predict what Actions are needed to achieve a desired Property/State of the adaptive glass.
Figure 3: A schematic representation of the architecture of MosyDB , with the concept of a ‘Box’ as a wrapper to a set of event nodes. (a) Generalized schema and topology of boxes and event nodes. (b) A schematic representation of the architecture of different types of boxes. An inventory box is a theme-related set of state nodes that describe individual states liked by a single theme, without a sequential topology. A protocol and experiment boxes are sequence-related event nodes, where the topology between the nodes is described by the relation between node ids (e.g., ), or by an ‘order’ attribute for a cyclic reaction-graphs. (c) A representation of an interconnected box emphasizing the inheritance property between boxes, leading to the possibility of branching out a main thread of a simple box via the source configuration attribute (see ‘configuration’ frame in (a)). Fig. 4 and Fig. 5 are examples that illustrate an implementation of these concepts.
Figure 4: Example of a reaction-graph, using the MosyDB architecture, of the following experiment: Student mixes 5 grams of material A with 7 grams of material B. To mix them, grin in a mortar and pestle. Presses the powder into a pellet with an Arbor press, and anneals under vacuum for 5 hours at 800 ∘ C. Collect diffractograms before and after annealing with a lab Cu x-ray source. . The additional linked boxes are a realization of the idea of interconnectivity between boxes.
Figure 5: Example on how using the MosyDB infrastructure, specifically an inventory-box, was used to collect property feature-list for data-driven extrapolation of materials and their properties. We demonstrate a use case in which the relationship among composition, electrical conductivity, and optical band gap is investigated. The additional linked boxes emphasize the capability to add reaction-graph information, which then can be used to build feature lists that are extracted from the nested box.
Figure 6: Decorated screen snapshots from the MosyDB webapp [ 19 ] , where (a) is the “My Boxes” window (after logging in), and (b) is the MosyDB -formatted content of the red box in (a). Each colored box in (a) corresponds to the boxes shown in (b): the main box (red) and two other ‘source box@id’ (blue and orange). The content of the red box and the related boxes (blue and orange) mimic the db-structured reaction-graph shown in Fig. 4 . One can find these boxes publicly available under the ‘Public’ tab in the ‘Home’ screen of the webapp. The QR codes in (a) contain the ‘box@id’, and can be used when linking via the ‘source box@id’ field in (b). Usage documentation can be found under the ‘Documentation’ tab of the webapp. The color-coded namespaces under each table resemble those presented in Fig. 3 (a). Note that the configurations ‘values’, ‘source,’ and ‘quantities’ are very common to almost any action. Configurations can also be added vertically under each event node (see in (b) - ‘order=#5’; ‘kind=Action’), where specialized configurations can be added. More comprehensive documentation for MosyDB is available at https://mosydb.odoo.com .
Figure 7: Similarity matrix of 83 MPG s from the Y-Mn-O dataset, relative to a base- MPG ( , marked ‘@’): (a) generated matrix, with experiments grouped by motivation; (b) experiments sorted by their summed similarity score; (c) reduced to the nodes in which quantity matters and re-sorted; (d) Pearson correlation between the powder x-ray diffraction pattern of the base experiment and those of all other experiments, sorted in descending order.
Figure 8: Large-scale similarity matrix, analogous to Fig. 7 (a), of about 1,300 hydrothermal zeolite syntheses. Blue indicates high similarity to the base- MPG , light colors low similarity, and dark red missing nodes (score of ‘-1’).