物質計測における機械学習応用と知識発見/Machine learning application and knowledge discovery in materials characterization

෺࣭ܭଌʹ͓͚ΔػցֶशԠ༻ͱ஌ࣝൃݟ   ୈ4ճ౷ܭɾػցֶशएखγϯϙδ΢Ϝ ໊ݹ԰޻ۀେֶ ླ໦༤ଠ ૯߹ݚڀେֶӃେֶߴΤωϧΪʔՃ଎ثݚڀػߏɹ෺࣭ߏ଄Պֶݚڀॴ

About me • ླ໦༤ଠ ʢYuta SUZUKIʣ https://resnant.github.io • Research Interest
• ࡐྉܭଌʹ޲͚ͨػցֶशͷԠ༻ • ओʹɺXઢΛ༻͍ͨແػ෺࣭ͷܭଌʢ࣓ੴɺి஑ͳͲʣ • ػցֶशΛ༻͍ͨ஌ࣝൃݟ • Education • ૯߹ݚڀେֶӃେֶɹߴΤωϧΪʔՃ଎ثՊֶݚڀՊ • খ໺ݚڀࣨ (PhD, 2019.3 - present) • ౦ژཧՊେֶɹجૅ޻ֶݚڀՊ • খ࢚ݚڀࣨ (BS & MS in Engineering) • Collaborators • ೔໺ӳҳ ઌੜʢ౷ܭ਺ཧݚڀॴʣ • ڇٱ঵޹ ઌੜʢΦϜϩϯαΠχοΫΤοΫεʣ • Funded by JST ACT-I & JSPS DC1 

,&,ߴΤωϧΪʔՃ଎ثݚڀػߏ • ೔ຊ࠷େͷ෺ཧͷݚڀॴʢ෺ཧతʹ΋େ͖͍ʣɿ500ਓҎ্ͷݚڀऀ • Ӊ஦ɺૉཻࢠ͔Βࡐྉɺੜ෺·Ͱ෯޿͍ݚڀ͕͞Ε͍ͯΔ • զʑ͸Xઢʢ์ࣹޫʣΛ࢖ͬͯ෺࣭Λௐ΂ΔݚڀʹऔΓ૊ΜͰ͍·͢ • ์ࣹޫɿ΄΅ޫ଎Ͱӡಈ͢Δిࢠ͔Β์ࣹ͞ΕΔXઢ
,&,

࣮ݧ૷ஔͳͲ

l.BUFSJBMT*OGPSNBUJDTzʢ#JP $IFNʹଓ͍ͯʣ w .BUFSJBMT*OGPSNBUJDT w ౷ܭతਪ࿦ʢػցֶशͳͲʣ΍৘ใॲཧٕज़ͷԠ༻ʹΑΓ  ޮ཰తͳࡐྉ։ൃΛ໨ࢦ֓͢೦ w ೥ʹถࠃ͕.(*ϓϩδΣΫτΛ։࢝ͯ͠Ҏདྷɺٸܹʹൃల
Perkins, K. et al. The 2019 materials by design roadmap. Journal of Physics D: Applied Physics 52, 013001 (2018).

.BUFSJBMT*OGPSNBUJDTͷجຊతํ๏࿦ γϛϡϨʔγϣϯͰσʔληοτ࡞੒ ػցֶशͰ෺ੑ༧ଌϞσϧΛߏங ϋΠεϧʔϓοτ࣮ݧͰݕূ ಘͨ஌ࣝΛʹϑΟʔυόοΫ
Ren, F. et al. Science Advances 4, eaaq1566 (2018).

.BUFSJBMT*OGPSNBUJDTͷجຊతํ๏࿦ γϛϡϨʔγϣϯͰσʔληοτ࡞੒ ػցֶशͰ෺ੑ༧ଌϞσϧΛߏங ϋΠεϧʔϓοτ࣮ݧͰݕূ ಘͨ஌ࣝΛʹϑΟʔυόοΫ
Ren, F. et al. Science Advances 4, eaaq1566 (2018). w ࡐྉ߹੒ w ଌఆ w ஌ࣝநग़ʢσʔλղੳʣ զʑ͸ɺ  ࣮ݧʢಛʹଌఆͱղੳʣͷޮ཰ԽΛ௨ͯ͡  .*ͷݚڀʹऔΓ૊ΜͰ͍·͢ɻ

.BUFSJBMT*OGPSNBUJDTݚڀಈ޲ • ࠃ಺֎ͷϓϩδΣΫτ • Materials Genome Initiative ʢถࠃ 2011ʙʣ
• NOMADʢEU 2015ʙʣɹ΄͔εΠεɺதࠃɺؖࠃͳͲ • MI2Iʢຊࠃ 2015ʙʣ • CREST ʰֵ৽ࡐྉ։ൃʱྖҬʢ2017ʙʣ • ͖͕͚͞ ʰ৘ใܭଌʱྖҬ ʢ2016ʙʣ • δϟʔφϧɺࠃࡍձٞ • NeurIPS Workshop (2017, 2018):  Machine Learning for Molecules and Materials • npj Computational Materialsʢ2015ʙ, IF = 9.651ʣˡMIͷઐ໳ࢽ • ͦͷଞ • Kaggle NOMAD ίϯϖʢಁ໌ಋిମͷ෺ੑ༧ଌίϯςετ, 2018೥ʣ • 800νʔϜҎ্͕ࢀՃɺ༏উऀ͸೔ຊਓ • N-gramʢࣗવݴޠॲཧ޲͚ಛ௃ྔʣΛ࢖͍ಛ௃ྔΛ࡞ͬͯ༏উ • ࠷ۙ͸Խ߹෺ͷ࣓ؾ૬ޓ࡞༻ͷ༧ଌίϯϖ΋ Sutton, C. et al. Crowd-sourcing materials-science challenges with the NOMAD 2018 Kaggle competition.   npj Comput. Mater. 5, 111 (2019).

1. σʔλ͕εϞʔϧɾߴ͍औಘίετ 2. WhyʁΛ໰ΘΕΔέʔε͕ଟ͍ 3. ໰୊ઃఆͷ೉͠͞ ࡐྉՊֶʹ͓͚Δػցֶशͷࠔ೉ ʢࡐྉܭଌͷཱ৔͔Βʣ

1. σʔλ͕εϞʔϧɾߴ͍औಘίετ • Φʔϓϯσʔλͷ੔උ͸·ͩಓ൒͹ • Materials Projectʢถࠃʣ • MatNaviʢ෺࣭ࡐྉݚڀػߏʣͳͲऔΓ૊Έ͕ਐߦத •
EUͰ͸ɺެతݚڀ݁ՌͷΦʔϓϯσʔλԽͷ࿮૊Έ͕੔උத • γϛϡϨʔγϣϯͰσʔλΛ४උɺ࣮ݧͱ૊Έ߹Θ͕ͤݱ࣮త • ࣮ࡍͷ࣮ݧͱͷΪϟοϓ͕՝୊ͱͳΔ • ࿹ͷݟͤॴͰ΋͋Δ • NLPΛ༻͍ͨจݙαʔϕΠͰσʔλऩू͢ΔࢼΈ΋͋Δ • σʔλͷϞμϦςΟ͕ଟ༷ • ςʔϒϧɺܥྻɺը૾ͳͲ͕ࠞࡏ ࡐྉՊֶʹ͓͚Δػցֶशͷࠔ೉ ʢࡐྉܭଌͷཱ৔͔Βʣ

1. σʔλ͕εϞʔϧɾߴ͍औಘίετ • ίετײͷࢀߟʢ͋͘·ͰҰྫͰ͕͢…ʣ • γϛϡϨʔγϣϯɿ1݅͋ͨΓ਺ඵʙ਺࣌ؒ • ࣮ݧɿ1݅͋ͨΓ1೔ʢ߹੒+ଌఆʣ • ࢼྉ1ݸ͋ͨΓ1000ԁʙ10ສԁͷՁ֨ײʢݪྉ͕ߴ͍ʣ
• ࣗಈԽɾޮ཰Խ΋ਐΈͭͭ͋Δ ࡐྉՊֶʹ͓͚Δػցֶशͷࠔ೉ ʢࡐྉܭଌͷཱ৔͔Βʣ ݸͰ૊੒ ݸͰ਺ඦ૊੒ w ૊੒܏ࣼബບ w ຕͰ૊੒Λ໢ཏͰ͖Δ

2. WhyʁΛ໰ΘΕΔέʔε͕ଟ͍ • ࡐྉՊֶऀ͸ɺ୯ʹԿ͔Λ༧ଌ͢Δ͚ͩͰ͸ͳ͘ɺ  ͦͷ༧ଌͷഎޙʹ͋Δ࢓૊Έʢ=෺ཧʣΛ஌Γ͍ͨ • ʮۚͷཛΛੜΉܲͦͷ΋ͷΑΓɺͦͷܲͷ࢓૊Έʹڵຯ͕͋Δʯ • ࠷ۙͷ࿦จͰ͸ɺ  MLʹΑΔ༧ଌ
+ ϞσϧͷղੳʹΑΔ஌ࣝൃݟ ͷߏ੒͕ଟ͍  • ༧ଌ݁ՌΛແ৚݅ʹड͚ೖΕΒΕΔ΄ͲMLʹ਌͠ΜͰ͍ͳ͍ • ೲಘײ͕ͳ͍ͱ࢖͍ͮΒ͍ ࡐྉՊֶʹ͓͚Δػցֶशͷࠔ೉ ʢࡐྉܭଌͷཱ৔͔Βʣ

3. ໰୊ઃఆͷ೉͠͞ • ͦ΋ͦ΋ɺͲΜͳ໰୊͕ػցֶशͰղ͚͏Δ͔ݕ౼͍ͯ͠Δஈ֊ • ղ͘΂͖໰୊Λݟ͚ͭग़͢εΩϧʢʁʣ͕͔ͳΓॏཁ • ΞΠσΟΞॏࢹͷ࿦จ͕ଟ͍  • ՝୊͕֤࿦తͰɺීวతͳλεΫ͕গͳ͍
• ςΫχοΫͷڞ༗΍ɺڝ૪͕೉͍͠ • ը૾ೝࣝίϯϖͳͲͱ͸ঢ়گ͕ҟͳ͍ͬͯΔ • ੒ޭͷࢦඪʢείΞ΍KPIʣ͕͸͖ͬΓ͠ͳ͍ • ఏҊख๏͸ɺ࣮ੈքͰ࣮ݧͯ͠͏·͍͔͘͘ௐ΂Δ͔͠ͳ͍ • ࣮ݧ͕େมͳܥͩͱ͔ͳΓͭΒ͍ • δϨϯϚɿେมͳ࣮ݧ΄Ͳޮ཰Խͷҙຯ͕͋Δ ࡐྉՊֶʹ͓͚Δػցֶशͷࠔ೉ ʢࡐྉܭଌͷཱ৔͔Βʣ

ࡐྉଌఆσʔλͷྫ 9ઢճંύλʔϯ  ݁থߏ଄ɺ૊੒ 9ઢٵऩεϖΫτϧ ిࢠঢ়ଶɺہॴߏ଄ 9ઢܬޫεϖΫτϧ ૊੒

ࡐྉଌఆσʔλͷྫ w ͍ͣΕ΋ɺੜσʔλΛղੳͯ͠ಘΒΕΔ෺ཧྔ͕ॏཁ 9ઢճંύλʔϯ  ݁থߏ଄ɺ૊੒ 9ઢٵऩεϖΫτϧ ిࢠঢ়ଶɺہॴߏ଄ 9ઢܬޫεϖΫτϧ ૊੒
• ੜσʔλ͸ෳࡶɺݟͯ΋Α͘Θ͔Βͳ͍ • σʔλղੳʹ͸ϞσϧϑΟοςΟϯάΛߦ͏͜ͱ͕ଟ͍ • ࢼߦࡨޡ͕ඞཁɺ͕͔͔࣌ؒΔʢʙ1೔/dataʣ

X-ray sample Detectors ͍͋ͪγϯΫϩτϩϯ BL5A (powder XRD beamline) σʔλͷ௡೾
ݱࡏͷݚڀࢪઃͰ͸ຖ೔100GBҎ্ͷσʔλ͕ੜΈग़͞ΕΔ • ྫɿ 5000ຊͷXRD͕24࣌ؒͰऔಘ͞ΕΔ (AichiSR) ͜ΕΒΛखಈͰղੳͰ͖ΔͩΖ͏͔ʁ

• Research purpose • ػցֶशΛ༻͍ͯɺߴ଎͔ͭࣗಈͷσʔλղੳٕज़Λ։ൃ͢Δ • έʔεελσΟ: 1. Xઢٵऩ෼ޫ๏ (XAS:
X-ray Absorption Spectroscopy) • σʔλͷੑ࣭Λௐ΂ͨݚڀ  ྨࣅ౓Λௐ΂Δ͚ͩͰ෺ཧྔ͕Θ͔ͬͨ 2. Xઢճં๏ (XRD: X-ray Diffraction) • ڭࢣ͋ΓֶशͷԠ༻ɺܾఆ໦Λؤுͬͯௐ΂ͯ஌ࣝൃݟ Outline

σʔλͷྨࣅ౓ʹண໨ͨ͠  9ઢٵऩεϖΫτϧ͔Βͷ෺ཧྔਪఆ έʔεελσΟ : 4V[VLJ et alMicroscopy and Microanalysis
: 4V[VLJ et alnpj Computational Materials,

Photon Energy (eV) w ෺࣭ʹ9ઢΛೖࣹ͠ɺͦͷٵऩεϖΫτϧΛௐ΂Δجຊతͳख๏ w ࡐྉͷ੒෼΍Խֶ݁߹ɺ࣓ؾঢ়ଶͳͲ͕Θ͔Δ 9ઢٵऩ෼ޫ๏ ,&,1IPUPO'BDUPSZ#-

2ͲΕʹࣅ͍ͯ·͔͢ʁ 9ઢٵऩ෼ޫ๏ ࢎԽϚϯΨϯʢ࣮ଌʣ γϛϡϨʔγϣϯ γϛϡϨʔγϣϯ γϛϡϨʔγϣϯ Mn 2+,  10Dq =
0.9 eV Mn 3+,  10Dq = 0.9 eV Mn 2+,  10Dq = 1.3 eV

w ෺࣭ʹ9ઢΛೖࣹ͠ɺͦͷٵऩεϖΫτϧΛௐ΂Δجຊతͳख๏ w ࡐྉͷ੒෼΍Խֶ݁߹ɺ࣓ؾঢ়ଶͳͲ͕Θ͔Δ w యܕతʹ͸ɺଌఆͨ͠εϖΫτϧΛจݙ΍ܭࢉ౳ͱݟൺ΂ɺ  ࣅͨ΋ͷΛ୳͢͜ͱͰ෺ཧྔΛਪఆʢσʔλղੳʣ͍ͯ͠Δ w ਓྗͷ࠷ۙ๣๏ w
໰୊఺ w ଌఆ͸ສ݅Iɺ  ͔݅͠͠ͷσʔλղੳʹ਺࣌ؒʙ਺೔Ҏ্ w ղੳऀͷओ؍͕আ͚ͳ͍ 9ઢٵऩ෼ޫ๏ EF(SPPU 'FUBM 1IZT3FW# r EF(SPPU ',PUBOJ "$PSF-FWFM4QFDUSPTDPQZPG4PMJET $3$1SFTT

• ຊ࣭తʹ͸ɺεϖΫτϧͷྨࣅ౓ΛଌΔई౓͕ͳ͍͜ͱ͕໰୊ • զʑͷΞΠσΟΞ:  దͨ͠ڑ཭ई౓Λ࢖͑͹εϖΫτϧͷྨࣅ౓ΛධՁͰ͖ͦ͏ • σʔλͷڑ཭ΛଌΔͱ͍͏ൃ૝͸ɺ  ࡐྉՊֶऀʹͱͯ͠͸ͳ͔ͳ͔ग़ͯ͜ͳ͔ͬͨ • ଟ༷ମԾઆ͔Βண૝
• ԾఆΛಋೖ • ʮࣅͨੑ࣭ͷ෺࣭ͷεϖΫτϧ͸ࣅ͍ͯΔ͸ͣʯ • దͨ͠ڑ཭ई౓ͱ͸ʁ • Euclidian distance, cosine, JS divergence… • ޙͰٞ࿦͠·͢ Key concepts

4XJTT3PMM ࣍ݩʹຒΊࠐ·Εͨ  ࣍ݩଟ༷ମ w ଟ༷ମԾઆ • “many data sets have
the property that the data points all lie close to a manifold of much lower dimensionality than that of the original data space” (Bishop 2006) w ʰߴ࣍ݩͳσʔλͰ͋ͬͯ΋ɺ  ɹ࣮ࡍʹ͸ΑΓ௿࣍ݩͳ෦෼ू߹ʹ෼෍͍ͯ͠Δʱ w ʮσʔλͷຊ࣭తͳ࣍ݩʢ࣮ޮ࣍ݩʣ͸௿͍  ʢ͜ͱ͕ଟ͍ʣʯ  w ༷ʑͳ࣍ݩ࡟ݮΞϧΰϦζϜ͕ߟҊ σʔλͷ࣮ޮ࣍ݩ *40."1 U4/& .%4

4XJTT3PMM ࣍ݩʹຒΊࠐ·Εͨ  ࣍ݩଟ༷ମ σʔλͷ࣮ޮ࣍ݩ w զʑͷΞΠσΟΞ w ܭଌσʔλͷ࣮ޮ࣍ݩ΋ɺ  ෺ཧతύϥϝʔλఔ౓ʢ਺ݸʣͷ࣍ݩ͔͠ͳ͍ͩΖ͏ w
ܭଌͱσʔλղੳͱ͸͢ͳΘͪɺ  ߴ࣍ݩσʔλ͔Β௿࣍ݩͷಛ௃ʢ෺ཧྔʣΛऔΓग़͢͜ͱɻ w ଟ༷ମֶशΛ༻͍ͨσʔλͷ࣍ݩ࡟ݮΛண૝ w ଟ༷ମԾઆ • “many data sets have the property that the data points all lie close to a manifold of much lower dimensionality than that of the original data space” (Bishop 2006) w ʰߴ࣍ݩͳσʔλͰ͋ͬͯ΋ɺ  ɹ࣮ࡍʹ͸ΑΓ௿࣍ݩͳ෦෼ू߹ʹ෼෍͍ͯ͠Δʱ w ʮσʔλͷຊ࣭తͳ࣍ݩʢ࣮ޮ࣍ݩʣ͸௿͍  ʢ͜ͱ͕ଟ͍ʣʯ

w ଟॏ߲͓Αͼ഑Ґࢠ৔ཧ࿦ʹج͖ͮ  .OQ9ઢٵऩεϖΫτϧΛγϛϡϨʔγϣϯ w Ձ਺ɿ w ݁থ৔ύϥϝʔλ%RɿʙF7
w ݁থ৔ͷରশੑ͸໘ମ഑Ґʹݻఆ w ߹ܭຊͷ9"4εϖΫτϧΛ४උ w εϖΫτϧຊ͸఺ͷσʔλϙΠϯτ w ࣍ݩ࡟ݮʹ͸ଟ࣍ݩई౓ߏ੒๏ʢ.%4ʣΛ༻͍ͨ w σʔλಉ࢜ͷڑ཭ΛͳΔ΂͘อͬͨ··ɺ  ΑΓ௿࣍ݩͳۭؒʹ͓͍ͯσʔλΛදݱ͢Δ w ࣍ݩˠ࣍ݩʹ࣍ݩ࡟ݮ σʔλͷ४උ Photon Energy (eV) 10 Dq (eV) Intensity ܭࢉͨ͠.O ͷ9"4 EF(SPPU ',PUBOJ "$PSF-FWFM4QFDUSPTDPQZPG4PMJET $3$1SFTT #PSH *(SPFOFO 1.PEFSONVMUJEJNFOTJPOBMTDBMJOHUIFPSZBOEBQQMJDBUJPOT 4QSJOHFS7FSMBH/FX:PSL

࣮ଌͨ͠.O0ͷεϖΫτϧ  ʢEF(SPPUFUBMʣ ਪఆ͞Εͨ෺ཧྔʹରԠ͢ΔεϖΫτϧ  ʢܭࢉɺ.O %RF7ʣ ैདྷ๏ͱ΋Α͘Ұக ݁Ռɿ࣍ݩ࡟ݮͱՄࢹԽ w εϖΫτϧಉ࢜ͷྨҎ౓͕  ෺ཧྔʹରԠͯ͠มԽͨ͠ 
ʢಋೖͨ͠Ծఆ͸ଥ౰ͩͬͨʂʣ w ྨࣅ౓͔Βਪఆ͞Εͨ෺ཧྔ͸  ख़࿅ऀʹΑΔಉఆʢैདྷ๏ʣͱҰக w σʔλͷྨࣅ౓ʹ΋ͱ͍ͮͯ  ෺ཧྔΛਪఆͰ͖Δͱࣔͤͨ ଟ࣍ݩई౓ߏ੒๏ʢ.%4ʣʹΑΔ࣍ݩ࡟ݮ݁Ռ ֤఺͕ຊͷ9"4εϖΫτϧɺ఺ಉ࢜ͷڑ཭͸εϖΫτϧͷ  ྨҎ౓ʹରԠ͢Δɻ৭͸Ձ਺ɺ਺ࣈ͸%RʢF7ʣYΛࣔ͢ ʢ͜͜Ͱ͸&VDMJEFBO%JTUBODFΛ༻͍͍ͯΔʣ ྨࣅ౓ʹج͍ͮͯɺ  ࣮ݧσʔλ͕ଥ౰ͳҐஔʹ  ϓϩοτ͞Ε͍ͯΔʢ੺ؙʣ

• ຊ࣭తʹ͸ɺεϖΫτϧͷྨࣅ౓ΛଌΔई౓͕ͳ͍͜ͱ͕໰୊ • զʑͷΞΠσΟΞ:  దͨ͠ڑ཭ई౓Λ࢖͑͹εϖΫτϧͷྨࣅ౓ΛධՁͰ͖ͦ͏ • ԾఆΛಋೖ • ʮࣅͨੑ࣭ͷ෺࣭ͷεϖΫτϧ͸ࣅ͍ͯΔ͸ͣʯ •
ଟ༷ମԾઆ͔Βண૝ • దͨ͠ڑ཭ई౓ͱ͸ʁ • ඞཁͳੑ࣭: • εϖΫτϧͷಛ௃Λଊ͑ΒΕΔ͜ͱ • ϐʔΫͷҐஔɺܗ… • ϊΠζͷ෇Ճʹ͍ͭͯϩόετͰ͋Δ͜ͱʢ࣮༻্ॏཁʣ • ϊΠζͱଌఆ࣌ؒ͸τϨʔυΦϑ Key conceptsʢ࠶ܝʣ

ϊΠζͷ෇Ճʹର͢Δݎ࿚ੑ w ϊΠζΛؚΜͩεϖΫτϧΛ࠶ݱ͢ΔͨΊɺ  ϊΠζΛΨ΢ε෼෍Ͱۙࣅ͠ɺεϖΫτϧʹՃࢉ w 4/ൺΛม͑ͯσʔληοτΛ࡞੒ w 4/ൺϐʔΫ࠷େ஋ϊΠζͷඪ४ภࠩͱఆٛ 9ઢٵऩεϖΫτϧ  ʢγϛϡϨʔγϣϯʣ
Ψ΢εϊΠζΛՃࢉ

ϊΠζͷ෇Ճʹର͢Δݎ࿚ੑ ϊΠζ෇Ճʹର͢Δڑ཭ؔ਺ͷݎ࿚ੑΛൺֱ w ϐΞιϯ૬ؔ܎਺͕ϕετɺDPTJOF΋ݎ࿚ w ϐʔΫͷڧ౓มԽʹରͯ͠ڑ཭͕ෆมͳͨΊͱਪ࡯͞ΕΔ w ϊΠδʔͳσʔλ͔ΒͰ΋෺ཧྔ͕ਪఆͰ͖Ε͹ɺ  ଌఆ࣌ؒΛେ෯ʹ୹ॖͰ͖Δ w
4/ൺ͕ͰΑ͚Ε͹ଌఆ࣌ؒ͸ͰࡁΉ ϊΠζେ ϦϑΝϨϯεʹର͢Δྨࣅ౓ ྑ͍

w εϖΫτϧͷྨҎ౓ΛଌΔ͚ͩͰɺͦͷࡐྉͷ෺ཧྔΛਪఆͰ͖ͨ w σʔλղੳΛେ෯ʹߴ଎ԽɺࣗಈԽͰ͖Δͱݟࠐ·ΕΔ  w ద੾ͳྨࣅੑई౓͕બఆ͞ΕͨʢϐΞιϯ૬ؔ܎਺͕ϕετʣ w ࡐྉσʔλ΁ͷػցֶशͷԠ༻ͷجૅͱͳΔ w ϊΠζ΍෼ղೳͷѱԽʹݎ࿚ͳ෺ཧྔਪఆ͕Մೳ
w ෆ҆ఆ෺࣭΍ߴ଎ݱ৅ͷଌఆͳͲɺ৽͍͠෺ཧ͕୓͚ΔՄೳੑ લ൒ͷ·ͱΊ

• Research purpose • ػցֶशΛ༻͍ͯɺߴ଎͔ͭࣗಈͷσʔλղੳٕज़Λ։ൃ͢Δ • έʔεελσΟ: 1. Xઢٵऩ෼ޫ๏ (XAS:
X-ray Absorption Spectroscopy) • σʔλͷੑ࣭Λௐ΂ͨݚڀ  ྨࣅ౓Λௐ΂Δ͚ͩͰ෺ཧྔ͕Θ͔ͬͨ 2. Xઢճં๏ (XRD: X-ray Diffraction) • ڭࢣ͋ΓֶशͷԠ༻ɺܾఆ໦Λؤுͬͯௐ΂ͯ஌ࣝൃݟ Outline

ػցֶशΛ༻͍ͨ  9ઢճંύλʔϯ͔Βͷ݁থߏ଄ਪఆ έʔεελσΟ : 4V[VLJet alMicroscopy and Microanalysis

ࡐྉ։ൃͷ࢝఺ɿ݁থߏ଄ղੳ • ෺࣭ͷػೳ͸݁থߏ଄ʹΑΓࢧ഑͞ΕΔ • ిؾ఻ಋੑɺ࣓ੑɺ৮ഔػೳɺFUDʜ • ݱ୅ͷࡐྉ։ൃͰ͸ݪࢠεέʔϧͰࡐྉΛ࡞ΓࠐΉ ݁থߏ଄ʢʹݪࢠͷฒͼํʣͷಉఆ͕ۃΊͯॏཁ • Ϧν΢ϜΠΦϯి஑ͷྫ
• ిۃࡐྉ-J$P0 ʹଞݩૉΛఴՃͯ݁͠থߏ଄Λ҆ఆԽ͠ɺ  ༰ྔྼԽΛ཈͑ͨిۃࡐྉ%-$0͕։ൃ Liu, Q. et al. Nature Energy (2018). %-$0ͷ9ઢճંύλʔϯ 9ઢճં๏ʹΑΔ݁থߏ଄ղੳ͸ࡐྉ։ൃͷ࠷ॏཁٕज़ͷҰͭ %-$0ͷ݁থߏ଄ "M -B -J

ࡐྉ։ൃͷ࢝఺ɿ݁থߏ଄ղੳ • ෺࣭ͷػೳ͸݁থߏ଄ʹΑΓࢧ഑͞ΕΔ • ిؾ఻ಋੑɺ࣓ੑɺ৮ഔػೳɺFUDʜ • ݱ୅ͷࡐྉ։ൃͰ͸ݪࢠεέʔϧͰࡐྉΛ࡞ΓࠐΉ ݁থߏ଄ʢʹݪࢠͷฒͼํʣͷಉఆ͕ۃΊͯॏཁ • Ϧν΢ϜΠΦϯి஑ͷྫ
• ిۃࡐྉ-J$P0 ʹଞݩૉΛఴՃͯ݁͠থߏ଄Λ҆ఆԽ͠ɺ  ༰ྔྼԽΛ཈͑ͨిۃࡐྉ%-$0͕։ൃ Liu, Q. et al. Nature Energy (2018). %-$0ͷ9ઢճંύλʔϯ 9ઢճં๏ʹΑΔ݁থߏ଄ղੳ͸ࡐྉ։ൃͷ࠷ॏཁٕज़ͷҰͭ %-$0ͷ݁থߏ଄ "M -B -J ੈքॳͷϦν΢ϜΠΦϯೋ࣍ి஑ͷిۃ  ʢࠓ೥ͷϊʔϕϧԽֶ৆ʹʂʣ

9ઢճંʢ93%ʣͷݪཧ • 9ઢճંʢ93%ʣ • ೾ͷڧΊ͍͋ɺऑΊ͍͋Λར༻ͯ݁͠থߏ଄Λௐ΂Δ • ϐʔΫͷҐஔ΍ߴ͞ͰɺݪࢠͷҐஔ΍छྨ͕Θ͔Δ • ϒϥοά৚݅ETJOВЕ
Xઢ ݕग़ث ֯౓Λม͑ͯଌఆ ճં֯ ճંڧ౓ Rigaku Corp.  https://www.rigaku.com/en/techniques/xrd

݁থߏ଄ղੳͷྲྀΕ • 9ઢճંʢ93%ʣ͸݁থߏ଄ղੳͷجຊख๏ͱͯ͠ීٴ • 93%ύλʔϯ͔Β݁থߏ଄͕Θ͔Δʢ೥ϊʔϕϧ৆ʣ • ෺࣭ݚڀɾ࢈ۀΛࢧ͑ΔෆՄܽͳπʔϧͷҰͭ • %/"ͷॏཐટߏ଄ͷൃݟʹ΋ߩݙʢ೥ϊʔϕϧ৆ʣ 93%ଌఆ૷ஔ
/E࣓ੴͷ93%ύλʔϯ /E࣓ੴͷ݁থߏ଄ ωΦδϜʢ/Eʣ࣓ੴ  ʢ/E 'F #ʣ 93%ύλʔϯͷ௚ײతཧղ͸೉͘͠ɺ  ݁থߏ଄ͷਪఆ͸ઐ໳Ոʹͱͬͯ΋༰қͰ͸ͳ͍ ˠޮ཰తͳσʔλղੳ͕՝୊ʹ ϐ ʔΫϑΟοςΟϯά * Chen, P. A. et al. J. Magn. Mag. Mater. 370, 45–53 (2014).

93%σʔλղੳͷ՝୊ • ଌఆσʔλͷղੳʹཁ͢Δ࣌ؒͱ࿑ྗ͕  ࡐྉݚڀͷϘτϧωοΫͱͳ͍ͬͯΔ • 93%ଌఆɿ݅EBZ • 93%ղੳɿ ݅EBZɹˠ͔͠΋σʔλͷେ൒͕ະ׆༻ •
ࣗಈंେख๭ࣾ͸ਓւઓज़ͰରԠ • 93%ղੳͷ೉͠͞ • શݪࢠͷ࣍ݩߏ଄͕ͭͷύλʔϯʹॏͳΓ߹͍ͬͯΔ • ղੳతʹ݁থߏ଄ΛٻΊΔ͜ͱ͸ෆՄೳʢෳૉ਺ͷҐ૬ΛଌఆͰ͖ͳ͍ͨΊʣ • طଘख๏ʢύϥ ϝʔλϑ Οος Οϯάʣ ͸ॳظ஋ܾఆ͕ඇৗʹ೉͍͠ ݁থߏ଄ 93%ύλʔϯ

Overview 1SFQBSJOH93%%BUBCBTF #VJMEJOH1SFEJDUJPO.PEFM ,OPXMFEHF%JTDPWFSZ &YQFSJNFOUBM7BMJEBUJPO

Overview 1SFQBSJOH93%%BUBCBTF #VJMEJOH1SFEJDUJPO.PEFM ,OPXMFEHF%JTDPWFSZ &YQFSJNFOUBM7BMJEBUJPO ͜ͷ෦෼ͷ࿩Λ͠·͢

0WFSWJFX 1SFQBSJOH93%%BUBCBTF #VJMEJOH1SFEJDUJPO.PEFM ,OPXMFEHF%JTDPWFSZ &YQFSJNFOUBM7BMJEBUJPO

݁থߏ଄༧ଌϞσϧߏங • ڭࢣ͋ΓֶशͷλεΫͱͯ͠ఆٛʢܗͱେ͖͞Λ༧ଌʣ ܗΛ༧ଌ େ͖͞Λ༧ଌ X-ray Diffraction Machine Learning
Pipeline Crystal System Prediction (Classification) Lattice Parameter Prediction (Regression) 1 ms Error < 0.8% Cubic, ɾ݁থܥʢΫϥεʣ  ɾۭؒ܈ʢΫϥεʣ

݁থߏ଄༧ଌϞσϧߏங • ڭࢣ͋ΓֶशͷλεΫͱͯ͠ఆٛʢܗͱେ͖͞Λ༧ଌʣ w࣮ݧʹద༻͢ΔͨΊɺ੍໿͋Γ w࣮ݧ৚͕݅ຖճมΘΔ w σʔλ௕ɺεΩϟϯ֯౓ൣғɺ೾௕ͳͲ w࣮ݧ৚݅ʹԠͯ͡ߴ଎ʹ࠶܇࿅͍ͨ͠ wࡶͳϋΠύϥͰ΋ͦͦ͜͜ಈ͍ͯ΄͍͠
wΞϧΰϦζϜʹ͸&YUSB5SFFTΛ༻͍ͨʢ3BOEPN'PSFTUͷѥछʣ w 3BOEPN'PSFTUΑΓ܇࿅͕਺ഒ଎͍ɺνϡʔχϯά΋؆୯ w ಛ௃ྔॏཁ౓ɺOEҎ߱ͷ༧ଌ݁Ռ΋ಘΒΕΔ wղੳ͠΍͍͢γϯϓϧͳಛ௃ྔΛબ୒ʢ࣍ݩʣ w ϐʔΫҐஔʢ௿֯ଆ͔Βຊʣɺ߹ܭϐʔΫຊ਺ X-ray Diffraction Machine Learning Pipeline Crystal System Prediction (Classification) Lattice Parameter Prediction (Regression) 1 ms Error < 0.8% Cubic,

݁Ռɿ݁থߏ଄ͷ༧ଌ݁Ռ ݁থܥͷࠞಉߦྻʢ7Ϋϥεʣ HgS a = 5.8717 Å CsCl a
= 4.1150 Å LiMn 2 O 4 a = 8.245 Å LiNi 0.5 Mn 2 Ti 1.2 O 4 a = 8.324 Å Hydrokenomicrolite a = 10.454 Å ֨ࢠఆ਺ͷ༧ଌ݁Ռ R2=0.9892 • ඇৗʹߴ͍ੑೳΛୡ੒ͨ͠ʢ͜Ε͸γϛϡϨʔγϣϯσʔλͰͷ݁Ռʣ • ෼ྨਫ਼౓ "DD ͸݁থܥɺۭؒ܈ɺ֨ࢠఆ਺ͷ3.4& • εΫϦʔχϯάʹ͸े෼ͳਫ਼౓ͱߟ͑Δ

ػցֶशϞσϧ͔Βͷ஌ࣝൃݟ • جຊతํ਑ • ݹయతΞϧΰϦζϜͰటष͘ؤுΔ • ख๏తʹ৽͍͠࿩͸ग़͖ͯ·ͤΜʢ͢Έ·ͤΜʣ • ࣮໰୊ͷྫͱͯ͠ࢀߟʹͳΕ͹޾͍Ͱ͢

݁ՌɿػցֶशϞσϧ͔Βͷ஌ࣝൃݟ • ண໨͢ΔϐʔΫຊ਺ʢಛ௃ྔͷ਺ʣͱɺ༧ଌਫ਼౓Λݕূ • ඞཁे෼ͳܭଌ৚݅ʢεΩϟϯ֯౓ʣͷܾఆ͕Ͱ͖Δ • ಘΒΕͨ஌ݟ • ʮ݁থܥ෼ྨͷͨΊʹண໨͢΂͖ϐʔΫ͸ࠨଆ͔Βຊఔ౓ͰΑ͍ʯ
• ೥Ҏ্࢖ΘΕ͍ͯΔܦݧଇʢϋφϫϧτ๏˞ʣͱ΋ྨࣅ  ˞ʮڧ౓ॱʹຊͷϐʔΫʹண໨ͤΑʯ ண໨͢ΔϐʔΫຊ਺ ݁থܥͷ༧ଌਫ਼౓ Q Q Qʜ XRDύλʔϯͷྫ

݁ՌɿػցֶशϞσϧ͔Βͷ஌ࣝൃݟ • ܇࿅ࡁΈͷ&YUSB5SFF͔Βಛ௃ྔͷॏཁ౓Λௐ΂ͨ • ܦݧଇͷ۩ମԽΛૂͬͨ • ݁থͷܗͷࣝผʹ͸ϐʔΫຊ਺ͷ৘ใ͕ॏཁɿׂͱࣗ໌ • ղऍɿʮϐʔΫͷຊ਺͸݁থͷରশੑʹରԠ͢Δʢফ໓ଇʣʯ
• ௿֯ଆͷϐʔΫҐஔʢQʙQ͋ͨΓʣ΋ॏཁͦ͏ • ෺ཧతͳղऍ͸ඇࣗ໌ ֤ಛ௃ͷॏཁ౓ ಛ௃ྔ ߹ܭϐʔΫຊ਺ QʢϐʔΫҐஔʣ Q Q Q Q Q Q Q Q Q Q Q Qʜ XRDύλʔϯͷྫ

݁ՌɿػցֶशϞσϧ͔Βͷ஌ࣝൃݟ • γϯϓϧͳܾఆ໦Λ܇࿅ɾՄࢹԽͯࣝ͠ผنଇΛݕূ ޙ೔ެ։༧ఆ

ఏҊख๏Λ࣮ݧతʹݕূ w ࢼྉɿ70ʢೋࢎԽόφδ΢Ϝʣ w ஝೤ࡐྉͱͯ͠஫໨͞ΕΔ࣍ੈ୅ࡐྉʮ෺Λˆʹอͭණʯ w ߏ଄૬సҠʹ൐͏જ೤Λར༻ w ಉ࣌ʹۚଐઈԑମసҠ͢ΔͨΊɺεΠονϯάૉࢠʹ΋ظ଴
w ిࢠͷεϐϯͱيಓͷடংԽʹΑΓ෺ੑ͕ൃݱ w ٯʹݴ͑͹ɺ݁থߏ଄Λௐ΂Ε͹ిࢠঢ়ଶ͕Θ͔Δ 67℃ ௿Թ૬ɿMonoclinicߏ଄  ʢઈԑମʙ൒ಋମʣ ߴԹ૬ɿTetragonalߏ଄ 

• ͍͋ͪγϯΫϩτϩϯʹͯ70ͷ93%ଌఆΛ࣮ࢪʢ#-4ʣ • ΄΅ޫ଎Ͱӡಈ͢Δిࢠ͔Β์ࣹ͞ΕΔେڧ౓9ઢΛར༻ XઢϏʔϜ ࢼྉ XઢΧϝϥ ์ࣹޫ࣮ݧͰݕূ༻σʔλΛऔಘ

• Թ౓มԽʹ൐͏70ͷߏ଄૬సҠ͕໌ྎʹ؍࡯͞Εͨ • 93%ຊ͸ඵͰࣗಈଌఆՄೳʢ࣌ؒͰຊʣ खಈͰͷղੳ݁ՌʢHSPVOEUSVUIʣ 70ͷ93%ύλʔϯʢ࣮ଌʣ ௿Թ૬ʢ31℃ʣ ௿Թ૬ʢ62℃ʣ ߴԹ૬ʢ64℃ʣ
૬సҠ

݁থܥ༧ଌ݁Ռ ۭؒ܈༧ଌ݁Ռʢ5PQʣ 139 [I4/mmm] 194 [P63/mmc] 198 [P213] 140
[I4/mcm] 136 [P42/mnm] Tetragonal : P42/mnm • ࣌ؒʢਓྗʣˠNTʢఏҊख๏ʣ • ݁থܥ͸ਖ਼ղʢΫϥεʣ • ۭؒ܈΋501Ͱਖ਼ղʢΫϥεʣ 70ʢߴԹ૬ʣͷ93%ύλʔϯʢ࣮ଌʣ ࣮σʔλΛ࢖ͬͯఏҊख๏Λݕূ • ࣮σʔλͰ΋࢖͑Δ͜ͱΛ֬ೝ

ޙ൒ͷ·ͱΊ • ػցֶशΛ༻͍9ઢճંύλʔϯ͔Β݁থߏ଄Λ༧ଌͨ͠ • εΫϦʔχϯάʹ͸े෼ͳਫ਼౓Λಘͨ • ࣮σʔλͰ΋͏·͘ಈ͍ͨ • ஌ࣝൃݟ
• ΍ͬͨ͜ͱ • &YUSB5SFFTͷಛ௃ྔॏཁ౓ͷղऍ • ܾఆ໦ʹΑΔۙࣅͱՄࢹԽɺղऍ • ख़࿅ऀͷצͷʢͷҰ෦ʣ۩ମԽ͕Ͱ͖ͨ • ܭଌޮ཰ԽͷͨΊͷࢦ਑Λಘͨ  • ݸਓతࡶײ • ܾఆ໦ϕʔεͷख๏͕ࢥͬͨΑΓڧྗͰڻ͍ͨ

ऴΘΓʹɿ.BUFSJBMT*OGPSNBUJDTݚڀಈ޲ • ࠓޙͷ.*ͷํ޲ੑʹ͍ͭͯ  ͭͷϙΠϯτΛࢦఠ 1. Share meaningful data •
σʔλ͸ܗࣜΛඪ४Խ͠ެ։͢΂͖ 2. Spur collaboration with competitions • େձͰڝ͑͹ٕज़ൃల͕Ճ଎͢Δ  ʢKaggle౳ʣ 3. Develop a shared language • ࡐྉͱػցֶशͷݚڀऀͷަྲྀΛ  ଅਐ͢΂͖ 4. Accelerate and automate Luna et al., Nature 552, 23-27 (2017)

ऴΘΓʹɿ.BUFSJBMT*OGPSNBUJDTݚڀಈ޲ • ࠓޙͷ.*ͷํ޲ੑʹ͍ͭͯ  ͭͷϙΠϯτΛࢦఠ 1. Share meaningful data •
σʔλ͸ܗࣜΛඪ४Խ͠ެ։͢΂͖ 2. Spur collaboration with competitions • େձͰڝ͑͹ٕज़ൃల͕Ճ଎͢Δ  ʢKaggle౳ʣ 3. Develop a shared language • ࡐྉͱػցֶशͷݚڀऀͷަྲྀΛ  ଅਐ͢΂͖ 4. Accelerate and automate Luna et al., Nature 552, 23-27 (2017) ݸਓతͳࡶײɿ • ๖ժظͳ͕ΒɺϓϨΠϠʔͷ૿Ճ͕ஶ͍͠ • ౎߹ͷ͍͍໰୊Λղ͚ͩ͘ʹݶΒͣɺ  ࣮໰୊΁ͷద༻ΛਐΊΔͷ͕࣍ͷεςοϓ • ͥͻίϥϘ͍ͤͯͩ͘͞͞ʂ

物質計測における機械学習応用と知識発見/Machine learning application and knowledge discovery in materials characterization

物質計測における機械学習応用と知識発見/Machine learning application and knowledge discovery in materials characterization

More Decks by Yuta Suzuki

Other Decks in Science

Featured

Transcript