{
  "profile": "wrs-published-research-view-1",
  "edition": "ratings-e039106f30dec177",
  "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS",
  "entry": {
    "id": "ai-oct-pinn-continuous-discrete",
    "title": "Physics-informed neural networks for PDE solution and parameter discovery",
    "summary": "Uses automatic differentiation of neural surrogate solutions to combine observed-data errors with differential-equation residual losses, supporting forward solution and inverse parameter fitting in related continuous/discrete-time formulations.",
    "boundary": "Two original 2017 parts bundled; not the invention of neural PDE solvers or automatic differentiation. Loss enforcement at finite collocation points does not certify exact global conservation, PDE satisfaction, identifiability, convergence or all-regime accuracy. Numerical and stochastic settings remain source-specific.",
    "area": "Physics-informed machine learning",
    "topic": "Artificial intelligence",
    "year": 2017,
    "lane": "HUMAN",
    "result": "ALGORITHM",
    "package_id": "ai-oct-pinn-continuous-discrete",
    "target": {
      "record_id": "urn:uuid:df22ab4c-d80a-416a-a631-8ba7f9839052",
      "version": 1,
      "content_hash": "sha256:2fa41f990b252aa3269634220c5f31cbfa40427637bc194392e70909518ade9e"
    },
    "humans": [
      "Maziar Raissi",
      "Paris Perdikaris",
      "George Em Karniadakis"
    ],
    "authors_complete": true,
    "human_role_note": "Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.",
    "agents": [],
    "related": [],
    "subresults": [],
    "scientific_assessment": {
      "source_review": "SOURCE_CURATED",
      "scientific_review": "NOT_REVIEWED_BY_WRS",
      "formal_replay": "NOT_RUN",
      "provider_attestation": "NONE"
    },
    "sources": [
      {
        "id": "src-ai-oct-1711-10561v1-primary",
        "title": "Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations",
        "url": "https://arxiv.org/pdf/1711.10561v1"
      },
      {
        "id": "src-ai-oct-1711-10566v1-primary",
        "title": "Physics Informed Deep Learning (Part II): Data-driven Discovery of Nonlinear Partial Differential Equations",
        "url": "https://arxiv.org/pdf/1711.10566v1"
      }
    ],
    "points": "320000",
    "published_grade": {
      "id": "ai-oct-pinn-continuous-discrete",
      "entry_id": "ai-oct-pinn-continuous-discrete",
      "band": "S",
      "modifier": "-",
      "target": {
        "record_id": "urn:uuid:df22ab4c-d80a-416a-a631-8ba7f9839052",
        "version": 1,
        "content_hash": "sha256:2fa41f990b252aa3269634220c5f31cbfa40427637bc194392e70909518ade9e"
      },
      "what_changed": "Uses automatic differentiation of neural surrogate solutions to combine observed-data errors with differential-equation residual losses, supporting forward solution and inverse parameter fitting in related continuous/discrete-time formulations.",
      "scope": "Uses automatic differentiation of neural surrogate solutions to combine observed-data errors with differential-equation residual losses, supporting forward solution and inverse parameter fitting in related continuous/discrete-time formulations. Two original 2017 parts bundled; not the invention of neural PDE solvers or automatic differentiation. Loss enforcement at finite collocation points does not certify exact global conservation, PDE satisfaction, identifiability, convergence or all-regime accuracy. Numerical and stochastic settings remain source-specific.",
      "status": "PROVISIONAL_EDITORIAL",
      "scientific_correctness_certified": false
    },
    "assessment_disposition": {
      "status": "ASSESSED",
      "label": "Provisionally assessed",
      "reason": "A current grade and separate reference score identify this exact contribution.",
      "target": {
        "record_id": "urn:uuid:df22ab4c-d80a-416a-a631-8ba7f9839052",
        "version": 1,
        "content_hash": "sha256:2fa41f990b252aa3269634220c5f31cbfa40427637bc194392e70909518ade9e"
      },
      "edition": "ratings-e039106f30dec177"
    }
  },
  "version": {
    "kind": "PUBLIC_RECORD_VIEW",
    "previous": null,
    "published_at": "2026-10-16T00:00:00Z",
    "publisher": {
      "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33",
      "role": "WRS_EDITORIAL_PUBLICATION"
    },
    "record": {
      "contributors": [
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Maziar Raissi",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations",
            "url": "https://arxiv.org/pdf/1711.10561v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Paris Perdikaris",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations",
            "url": "https://arxiv.org/pdf/1711.10561v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "George Em Karniadakis",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations",
            "url": "https://arxiv.org/pdf/1711.10561v1"
          }
        }
      ],
      "evidence": [
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-0",
          "kind": "source_reference",
          "source": {
            "citation": "Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations",
            "url": "https://arxiv.org/pdf/1711.10561v1"
          }
        },
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-1",
          "kind": "source_reference",
          "source": {
            "citation": "Physics Informed Deep Learning (Part II): Data-driven Discovery of Nonlinear Partial Differential Equations",
            "url": "https://arxiv.org/pdf/1711.10566v1"
          }
        }
      ],
      "origin": "HUMAN_LEGACY",
      "record_profile": "wrs-local-research-notes-0.1",
      "record_type": "research_note",
      "rrp_version": "0.1",
      "scope": {
        "description": "Two original 2017 parts bundled; not the invention of neural PDE solvers or automatic differentiation. Loss enforcement at finite collocation points does not certify exact global conservation, PDE satisfaction, identifiability, convergence or all-regime accuracy. Numerical and stochastic settings remain source-specific.",
        "domain": "Physics-informed machine learning"
      },
      "sources": [
        {
          "citation": "Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations",
          "url": "https://arxiv.org/pdf/1711.10561v1"
        },
        {
          "citation": "Physics Informed Deep Learning (Part II): Data-driven Discovery of Nonlinear Partial Differential Equations",
          "url": "https://arxiv.org/pdf/1711.10566v1"
        }
      ],
      "statement": "# Physics-informed neural networks for PDE solution and parameter discovery\n\n## The question\n\nA physical equation can tell us how a system should behave even when observations are scarce. Can that equation help train a neural approximation? Physics-informed neural networks combine agreement with observed data and agreement with differential-equation constraints. They can be used for forward prediction or for estimating unknown model parameters from measurements.\n\n## The idea\n\nA neural network represents a candidate solution. Automatic differentiation computes derivatives of that candidate, and the equation turns those derivatives into a residual: a number measuring how far the equation is from being satisfied at a tested location. Training reduces a combination of residual, observation and boundary-condition errors. The choice of points and relative loss weights matters.\n\n## Why it matters\n\nThe contribution is a broadly reusable integration of differentiable models, data fitting and equation constraints. Earlier neural approaches to differential equations retain credit. Calling a model physics-informed does not make every prediction physically correct. A small average training loss can coexist with missed regions, inaccurate boundary behavior, parameter ambiguity or unstable behavior outside the tested range.\n\n## A teaching example\n\nConsider an original classroom spring model. In a simple idealization, acceleration is proportional to minus displacement. A candidate position curve therefore implies both a displacement and a second derivative at each time. Their mismatch provides an equation residual. Checking that mismatch at a hundred selected times still does not prove it is small between those times, especially for a rapidly varying candidate curve.\n\n## Scope and a useful next step\n\nStart with an equation whose solution is already known and keep a separate, denser checking grid. Compare solution error with training residual rather than reporting only the latter. State initial and boundary conditions and test sensitivity to sampling locations. An informative failure can show where a low residual ceases to be a reliable guide to actual solution accuracy. Two original 2017 parts bundled; not the invention of neural PDE solvers or automatic differentiation. Loss enforcement at finite collocation points does not certify exact global conservation, PDE satisfaction, identifiability, convergence or all-regime accuracy. Numerical and stochastic settings remain source-specific.\n\n## Scientific scope\n\nTwo original 2017 parts bundled; not the invention of neural PDE solvers or automatic differentiation. Loss enforcement at finite collocation points does not certify exact global conservation, PDE satisfaction, identifiability, convergence or all-regime accuracy. Numerical and stochastic settings remain source-specific.\n\n## Sources for this explanation\n\n- [Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations](https://arxiv.org/pdf/1711.10561v1): Pinned original contribution and its stated method, assumptions and limits. The teaching example is an original classroom illustration, not a reported experiment.\n- [Physics Informed Deep Learning (Part II): Data-driven Discovery of Nonlinear Partial Differential Equations](https://arxiv.org/pdf/1711.10566v1): Pinned original contribution and its stated method, assumptions and limits. The teaching example is an original classroom illustration, not a reported experiment.\n",
      "summary": "Uses automatic differentiation of neural surrogate solutions to combine observed-data errors with differential-equation residual losses, supporting forward solution and inverse parameter fitting in related continuous/discrete-time formulations.",
      "title": "Physics-informed neural networks for PDE solution and parameter discovery"
    },
    "record_id": "urn:uuid:df22ab4c-d80a-416a-a631-8ba7f9839052",
    "version": 1,
    "source_ref": {
      "record_id": "urn:uuid:df22ab4c-d80a-416a-a631-8ba7f9839052",
      "version": 1,
      "content_hash": "sha256:2fa41f990b252aa3269634220c5f31cbfa40427637bc194392e70909518ade9e"
    },
    "export_view": "wrs-public-editorial-view-1",
    "public_view_hash": "sha256:c46176928d5e3ba84448d2958914d7a630d7b7b507694c032c5847060a6a0888"
  },
  "target": {
    "record_id": "urn:uuid:df22ab4c-d80a-416a-a631-8ba7f9839052",
    "version": 1,
    "content_hash": "sha256:2fa41f990b252aa3269634220c5f31cbfa40427637bc194392e70909518ade9e"
  },
  "published_grade": {
    "profile": "wrs-published-grade-1",
    "edition": "2026-10-28.1",
    "id": "ai-oct-pinn-continuous-discrete",
    "entry_id": "ai-oct-pinn-continuous-discrete",
    "revision": 1,
    "title": "Physics-informed neural networks for PDE solution and parameter discovery",
    "context": "Physics-informed machine learning",
    "track": "Human-designed research method or resource",
    "band": "S",
    "modifier": "-",
    "status": "PROVISIONAL_EDITORIAL",
    "assessed_on": "2026-10-16",
    "evidence_cutoff": "2026-10-16",
    "achievement_state": "SUPPORTED_SCOPED_RESULT",
    "what_changed": "Uses automatic differentiation of neural surrogate solutions to combine observed-data errors with differential-equation residual losses, supporting forward solution and inverse parameter fitting in related continuous/discrete-time formulations.",
    "scope": "Uses automatic differentiation of neural surrogate solutions to combine observed-data errors with differential-equation residual losses, supporting forward solution and inverse parameter fitting in related continuous/discrete-time formulations. Two original 2017 parts bundled; not the invention of neural PDE solvers or automatic differentiation. Loss enforcement at finite collocation points does not certify exact global conservation, PDE satisfaction, identifiability, convergence or all-regime accuracy. Numerical and stochastic settings remain source-specific.",
    "evidence_summary": "The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Two original 2017 parts bundled; not the invention of neural PDE solvers or automatic differentiation. Loss enforcement at finite collocation points does not certify exact global conservation, PDE satisfaction, identifiability, convergence or all-regime accuracy. Numerical and stochastic settings remain source-specific.",
    "references": [
      {
        "title": "Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations",
        "url": "https://arxiv.org/pdf/1711.10561v1",
        "locator": "Original contribution and stated assumptions, evaluation scope and limitations; source-linked assessment, not independent execution or proof certification."
      },
      {
        "title": "Physics Informed Deep Learning (Part II): Data-driven Discovery of Nonlinear Partial Differential Equations",
        "url": "https://arxiv.org/pdf/1711.10566v1",
        "locator": "Original contribution and stated assumptions, evaluation scope and limitations; source-linked assessment, not independent execution or proof certification."
      }
    ],
    "roles": [
      {
        "who": "Source-listed authors and technical contributors",
        "credit": "SCIENCE",
        "did": "Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities."
      }
    ],
    "target": {
      "record_id": "urn:uuid:df22ab4c-d80a-416a-a631-8ba7f9839052",
      "version": 1,
      "content_hash": "sha256:2fa41f990b252aa3269634220c5f31cbfa40427637bc194392e70909518ade9e"
    },
    "scientific_correctness_certified": false,
    "independent_expert_calibration": false,
    "method": "PROPRIETARY",
    "points": "320000"
  },
  "canonical_agent_submission": false,
  "scientific_verification_awarded": false,
  "reader_guide": {
    "id": "ai-oct-pinn-continuous-discrete",
    "target": {
      "record_id": "urn:uuid:df22ab4c-d80a-416a-a631-8ba7f9839052",
      "version": 1,
      "content_hash": "sha256:2fa41f990b252aa3269634220c5f31cbfa40427637bc194392e70909518ade9e"
    },
    "sections": [
      {
        "title": "The question",
        "text": "A physical equation can tell us how a system should behave even when observations are scarce. Can that equation help train a neural approximation? Physics-informed neural networks combine agreement with observed data and agreement with differential-equation constraints. They can be used for forward prediction or for estimating unknown model parameters from measurements."
      },
      {
        "title": "The idea",
        "text": "A neural network represents a candidate solution. Automatic differentiation computes derivatives of that candidate, and the equation turns those derivatives into a residual: a number measuring how far the equation is from being satisfied at a tested location. Training reduces a combination of residual, observation and boundary-condition errors. The choice of points and relative loss weights matters."
      },
      {
        "title": "Why it matters",
        "text": "The contribution is a broadly reusable integration of differentiable models, data fitting and equation constraints. Earlier neural approaches to differential equations retain credit. Calling a model physics-informed does not make every prediction physically correct. A small average training loss can coexist with missed regions, inaccurate boundary behavior, parameter ambiguity or unstable behavior outside the tested range."
      },
      {
        "title": "A teaching example",
        "text": "Consider an original classroom spring model. In a simple idealization, acceleration is proportional to minus displacement. A candidate position curve therefore implies both a displacement and a second derivative at each time. Their mismatch provides an equation residual. Checking that mismatch at a hundred selected times still does not prove it is small between those times, especially for a rapidly varying candidate curve."
      },
      {
        "title": "Scope and a useful next step",
        "text": "Start with an equation whose solution is already known and keep a separate, denser checking grid. Compare solution error with training residual rather than reporting only the latter. State initial and boundary conditions and test sensitivity to sampling locations. An informative failure can show where a low residual ceases to be a reliable guide to actual solution accuracy. Two original 2017 parts bundled; not the invention of neural PDE solvers or automatic differentiation. Loss enforcement at finite collocation points does not certify exact global conservation, PDE satisfaction, identifiability, convergence or all-regime accuracy. Numerical and stochastic settings remain source-specific."
      }
    ],
    "references": [
      {
        "title": "Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations",
        "url": "https://arxiv.org/pdf/1711.10561v1",
        "locator": "Pinned original contribution and its stated method, assumptions and limits. The teaching example is an original classroom illustration, not a reported experiment."
      },
      {
        "title": "Physics Informed Deep Learning (Part II): Data-driven Discovery of Nonlinear Partial Differential Equations",
        "url": "https://arxiv.org/pdf/1711.10566v1",
        "locator": "Pinned original contribution and its stated method, assumptions and limits. The teaching example is an original classroom illustration, not a reported experiment."
      }
    ]
  },
  "reading_status": "EXPLANATORY_GUIDE"
}