PaperlayerView exact revision on arXiv

Source-linked AI summary

Complex materials engineering via Provenance Action Graphs: potential, structure and demonstration

Yevgeny Rakita, Paul Todd, James R. Neilson, Simon J. L. Billinge

arXiv:2610.18314v1cs.AIcond-mat.mtrl-sci

TL;DR

Synthesis databases often lack structured sequence information, despite synthesis history helping distinguish material states and properties. This work implements MosyDB to capture actions, observations, and characterization in provenance graphs, and reports that recorded-history similarity is only a weak predictor of resulting-state similarity.

  • Problem

    Structured synthesis-history data are needed to connect processing pathways with material states and properties, while failures are rarely available in machine-readable form.

    Source: §I · p. 2, B. Data structure and internal… · p. 3

  • Method

    MosyDB records synthesis actions, outputs, and characterization results in action graphs and relates History, State, and Property.

    Source: §I · p. 1, B. Data structure and internal… · p. 3

  • Results

    Similarity in recorded history was only a weak predictor of resulting state: some high-history-score experiments had diffraction patterns essentially uncorrelated with the base (r < 0.1).

    Source: C. Interaction with the data · p. 10

  • Takeaways & Limitations

    The authors conclude that MosyDB's value grows with richer structured data and that sharing reusable boxes can reduce vocabulary overhead across communities.

    Source: §V · p. 12

    Reusing items through pick-and-drop speeds data entry but introduces a risk associated with copied items.

    Source: §IV · p. 10

Abstract

from arXiv · show

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that follows years of trial and error, culminating in (published) successful recipes, but is truly driven by far more (unpublished) failures.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.

I. INTRODUCTION

The introduction argues that synthesis history and its sequence are essential for understanding material states and engineering desired properties. It presents MosyDB as an action-graph database for capturing synthesis and exploring how structured records can support learning and collaboration.

I. INTRODUCTION

  • Synthesis history can determine material state beyond composition, with processing conditions and step order linked to different phases, defects, chemical environments, and properties.The authors emphasize solid-state synthesis and materials including battery electrodes, high-entropy alloys, catalysts, thermoelectrics, and photovoltaics.Source: §I · p. 1
  • A FAIR database of synthesis paths could help address the high-dimensional search for desired properties, where intuition may be inaccessible to novices or miss routes.The authors propose machine learning to augment experts, support novice training, or reverse-engineer synthesis from measured properties.Source: §I · p. 1
  • MosyDB captures synthesis inputs, outputs, and characterization results between actions, creating records that can relate pathways to outcomes.The authors describe learning which actions or parameters distinguish pathways and, ultimately, predicting protocols for target states or properties.Source: §I · pp. 1–2
  • A uniform, machine- and human-readable schema is proposed to integrate automated and discrete synthesis, record failures and successes, reduce language misinterpretation, and support cross-field communication.The authors identify agreement on both field-agnostic and field-specific schemas as a significant challenge.Source: §I · p. 2
  • MosyDB demonstrates capturing chemical syntheses as action graphs, while the paper explores challenges to realizing the approach across laboratories and disciplines.The authors describe it as a start toward a high-level schema, not an attempt to justify one particular schema.Source: §I · p. 2
  • Many synthesis databases treat history as unrestricted metadata, while MosyDB captures synthesis as actions or sequential events from which material state can later be derived.Source: §I · p. 2

A. The structure of a MPG

A materials provenance graph represents synthesis as linked actions and observations, with sequence encoded in edges and event-specific information carried as configurations. Its granularity can vary, allowing complex operations to be expanded or bundled, and MosyDB provides a prototype for entering such data.

A. The structure of a MPG

  • A synthesis can be represented as time-ordered action and observation events, or as a directed graph whose edges encode action sequence and whose branches can split and merge.The graph abstraction can be mapped back to the time-based representation.Source: A. The structure of a MPG · p. 2
  • Actions change the sample, while observations measure its state; actions function as inputs or features and observations as outputs or labels.Observations may include color, texture, images, spectra, or structural signatures at points in the sequence.Source: A. The structure of a MPG · p. 2
  • Action and observation nodes can carry configurations, including details such as added material and amount; notes can preserve free-written information when designated fields are difficult to guarantee.If apparently identical graphs yield different observations, the discrepancy can point to an uncaptured parameter or action and motivate schema evolution.Source: A. The structure of a MPG · pp. 2–3, Figure 1 · p. 3
  • MPG granularity is variable: a complex operation can be bundled as a subgraph or expanded to expose actions and attributes that matter to the outcome.This lets experimenters hide details that do not vary or are not critical.Source: A. The structure of a MPG · p. 3, Figure 1 · p. 3
  • MosyDB is a prototype portal through which users entered MPG data into an experimental database, intended to explore implementation and schema challenges with early adopters.The project considers top-level and field-specific schemas.Source: A. The structure of a MPG · p. 3

B. Data structure and internal relationships for rationalized decision making

This section frames rationalized materials decisions through a History → State → Property relationship: ordered actions shape the measured state, which informs property readouts. It proposes a sequence-aware matrix representation of synthesis history to support comparisons across experiments.

B. Data structure and internal relationships for rationalized decision making

  • A State-observable is a structural or chemical descriptor, whereas a Property-observable is a readout such as resistivity or dielectric constant used for engineering decisions.The distinction is framed around technology-relevant prediction.Source: B. Data structure and internal… · p. 3
  • The proposed History → State → Property relationship treats history as determining state and state as informing property, supporting engineering decisions based on known material state.The authors contrast this with a cook-and-look approach that requires reproducing the full set of actions.Source: B. Data structure and internal… · pp. 3–4
  • Each experiment is defined as a set of actions between adjacent observables, providing a common dimension for cross-correlations.State sets may contain multidimensional measurements such as diffraction patterns or spectroscopy.Source: B. Data structure and internal… · p. 4
  • Synthesis history is represented as a 2D matrix with sequence on one axis and values uniquely linked to actions or attributes on the other.The example encodes adding Pb-acetate and KI, heating at 55 C, and waiting 60 minutes as ordered steps.Source: B. Data structure and internal… · p. 4
  • The matrix can distinguish sequence-important actions from attributes and can represent unspecified values with a consistent unknown marker.The examples make heat sequence-important while treating wait as its attribute, and use -1 for an unspecified temperature.Source: B. Data structure and internal… · p. 4

C. The architecture of MosyDB

MosyDB uses a field-agnostic schema to represent synthesis histories as connected actions, observations, and configurations. Its boxes organize these records by theme or sequence and support their reuse across workflows.

C. The architecture of MosyDB

  • A field-agnostic top-level schema is intended to let different fields and labs share machine-interpretable synthesis knowledge.Field-specific keyword choices can hinder communication between fields.Source: C. The architecture of MosyDB · pp. 4–5
  • Event nodes use shared fields for identity, kind, domain, name, description, and configurations, with IDs or order attributes encoding event sequence.Configurations hold synthesis-specific values such as material quantity, heating rate, or instrument settings.Source: C. The architecture of MosyDB · p. 5
  • Boxes can reference other boxes, enabling nested reuse and inheritance that can branch a main workflow through source configurations.A protocol becomes an experiment once executed.Source: Figure 3 · p. 6, C. The architecture of MosyDB · p. 6
  • Reusing shared boxes may help communities converge on a limited vocabulary and reduce barriers to a global machine- and human-readable infrastructure.Nodes, graphs, and configurations can be reused between boxes.Source: C. The architecture of MosyDB · pp. 5–6
  • A box bundles event nodes and configurations under a unique ID, with its type defining whether contents are organized by theme or sequence.Inventory boxes are thematic; protocol and experiment boxes contain sequenced events, and experiments require recorded observations.Source: C. The architecture of MosyDB · p. 6, Figure 3 · p. 6

A. Use-cases

MosyDB is demonstrated with a sequential solid-state synthesis graph and with an inventory-box representation of a property feature table. Linking samples to experiment boxes lets history-derived features augment the table without significant restructuring.

A. Use-cases

  • The two use cases demonstrate both a typical synthesis experiment and the mapping of state-oriented feature-label data into MosyDB with linked reaction graphs.The inventory representation stores sample rows as feature-list observation nodes and table columns as configurations.Source: A. Use-cases · pp. 6–7
  • A solid-state synthesis graph records adding and mixing materials, pressing a pellet, measuring XRD, annealing, and measuring XRD again.Configuration-based queries can filter matching sequences by conditions such as annealing temperature, material quantity, or press type, and compare measurements before and after annealing.Source: A. Use-cases · pp. 6–7, Figure 4 · p. 7
  • Linked experiment histories let MosyDB append annealing temperature and duration to property feature tables, supporting pathway-property correlations without significant restructuring.The demonstrated inventory-box use case investigates composition, electrical conductivity, and optical band gap.Source: A. Use-cases · pp. 7–8, Figure 5 · p. 7

B. User Interface

MosyDB’s web interface organizes synthesis records as reusable boxes containing connected event nodes and configuration items. Its order attribute simplifies representing reaction sequences, while the interface also supports sharing and QR-linked references.

B. User Interface

  • MosyDB’s interface combines a relational-cloud backend with a web app, plus early Python tools for bulk entry, analysis, filtering, and visualization.The paper focuses on the web interface, which works on PCs and mobile devices.Source: B. User Interface · p. 8
  • Users create typed boxes with metadata and unique IDs; boxes can be duplicated without content and populated with nodes and configurations.QR codes provide quick references to box IDs, and metadata supports filtering and identification.Source: B. User Interface · p. 8
  • Each line in a box represents an event node or configuration, with common fields and options to add specialized configurations or link configuration boxes.The interface supports reuse of nodes and configurations to increase reproducibility.Source: B. User Interface · p. 8
  • An order attribute represents node connectivity in a spreadsheet-friendly sequence, but the current implementation does not support cyclic reaction graphs.Explicit topology could represent cyclic operations; inherited boxes can represent branching graphs.Source: B. User Interface · p. 8
  • QR codes can link materials or instruments to boxes, supporting direct scanning into the source box field.The figure shows a main box and related source boxes whose contents mimic a database-structured reaction graph.Source: B. User Interface · pp. 8–9, Figure 6 · p. 9

C. Interaction with the data

The authors compare synthesis histories across 83 Y-Mn-O experiments using a similarity matrix, then compare those histories with diffraction patterns. The comparisons show that recorded-history similarity can diverge from similarity in the resulting state.

C. Interaction with the data

  • The similarity matrix compared the base experiment’s action-node domain, name, and quantity against 83 experiments, with rows for base actions and columns for experiments.The base-MPG ranked first with a summed score of 11; auxiliary boxes describing different operations ranked last.Source: C. Interaction with the data · p. 10
  • Filtering to chemically meaningful quantities—precursor amounts and annealing temperature—produced a re-sorted history comparison.The initial base-MPG was the experiment aimed at purifying o-YMnO3.Source: C. Interaction with the data · p. 10
  • Similar recorded histories were only a weak predictor of similar material states: some experiments with high history scores had diffraction correlations below 0.1, while other routes reached highly correlated structures (r > 0.95).The authors suggest missing parameters, such as reaction atmosphere or precursor identity, may explain discrepancies and that combining history- and state-based similarity can help identify them.Source: C. Interaction with the data · p. 10
  • A similar map was generated for about 1,300 hydrothermal zeolite syntheses, where families of related experiments and graph lengths could be seen at a glance.Dark-red regions indicate missing nodes.Source: C. Interaction with the data · p. 10

IV. DISCUSSION

The discussion identifies data-entry burden and weak immediate incentives as practical barriers to adoption. It also argues that well-structured, complete synthesis histories are valuable for analysis, while noting that the demonstrated datasets remain limited in scale and breadth.

IV. DISCUSSION

  • Recording a material provenance graph was an additional task because experiments are typically planned and recorded at the bench, often in notebooks or not at all.Structured entry was usually performed separately from the experiment and afterward.Source: §IV · p. 10
  • The cost of entering data falls immediately on individual users, while the shared database’s value grows only as its dataset becomes richer.The authors believe adoption depends more on faster, easier entry and clear rewards—such as analysis, comparisons, or recommendations—than on schema details.Source: §IV · p. 10
  • Reusing copied items can save data-entry effort but risks carrying values that do not match the experiment performed.The paper points to direct bench capture as a way to reduce both effort and this source of error.Source: §IV · pp. 10–11
  • The paper presents a similarity matrix for about 1,300 hydrothermal zeolite syntheses, illustrating that the map can be generated at larger scale.The figure encodes high similarity in blue, low similarity in light colors, and missing nodes in dark red.Source: Figure 8 · p. 11
  • Bench-side capture could make the graph a by-product of experimentation, and augmented-reality lab tools could extend this approach.The authors envision an assistant recording actions and observations as they occur and transferring them into the graph.Source: §IV · pp. 11–12
  • Although MosyDB held more than 2,500 experiments and protocols, the analyses used datasets of tens to about a thousand MPGs from a small number of groups.Larger datasets spanning more groups and fields are needed to establish whether the representation delivers the predictive power that motivates it.Source: §IV · p. 12

V. CONCLUSIONS

The conclusion presents MosyDB as an implementation of material provenance graphs and emphasizes shared, structured boxes as a basis for accumulating and translating synthesis knowledge.

V. CONCLUSIONS

  • The paper discusses the material provenance graph concept and demonstrates its implementation through MosyDB.The conclusion describes boxes as the currency of MosyDB, gaining value as structured data accumulates.Source: §V · p. 12
  • Sharing and reusing boxes can help communities converge on a narrower vocabulary while allowing scientists to add relevant attributes.Users can search for configured nodes, place them in a new experiment or protocol box, and share the result locally or globally.Source: §V · p. 12
  • A converging vocabulary is expected to make the data more machine-interpretable and reduce interoperability overhead across communities.Source: §V · p. 12
Loading 2610.18314v1…